Datasets:
benchmark_id stringlengths 41 44 | source stringclasses 2
values | source_record_id stringlengths 25 77 | source_snapshot_at stringdate 2026-08-10 02:44:31 2026-08-10 02:44:31 ⌀ | measured_at stringlengths 20 24 | schema_version int64 1 1 | model_id stringlengths 7 65 | model_revision stringclasses 9
values | model_display_name stringclasses 176
values | base_model_id stringclasses 96
values | model_family stringclasses 13
values | model_params_billion float64 0 304 ⌀ | model_active_params_billion int64 1 4 ⌀ | model_is_moe bool 2
classes | runtime stringclasses 9
values | runtime_version stringclasses 106
values | backend stringclasses 5
values | quantization stringclasses 67
values | hardware_label stringclasses 39
values | hardware_class stringclasses 3
values | accelerator stringclasses 30
values | accelerator_count int64 1 8 | vram_gb float64 0.5 383 ⌀ | unified_memory_gb int64 24 128 ⌀ | ram_gb float64 16 377 ⌀ | cpu stringclasses 41
values | cpu_cores int64 4 384 ⌀ | os stringclasses 21
values | prompt_tokens int64 0 256k ⌀ | output_tokens float64 0 6k | context_length int64 512 1.05M ⌀ | batch_size int64 1 8.19k ⌀ | num_runs int64 3 10 ⌀ | ttft_ms float64 0.5 164k ⌀ | output_tps float64 0.88 4.98k | prefill_tps float64 1 270k ⌀ | total_tps float64 4.3 110k ⌀ | min_tps float64 1.31 785 ⌀ | max_tps float64 1.45 791 ⌀ | total_time_ms float64 373 218k ⌀ | peak_vram_gb float64 0.01 23.7k ⌀ | gpu_power_watts listlengths 0 2 | total_power_watts float64 61.8 685 ⌀ | tensor_parallel int64 1 4 ⌀ | gpu_layers int64 -1 1,000k ⌀ | kv_cache_dtype stringclasses 14
values | attention_backend stringclasses 6
values | flash_attention bool 2
classes | speculative_decoding bool 2
classes | mtp_enabled bool 2
classes | submitter stringclasses 71
values | submitter_verified bool 1
class | notes stringlengths 17 1.98k ⌀ |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
external-community:cmoetb4n1000jjz04aol3g3h5 | external-community | cmoetb4n1000jjz04aol3g3h5 | 2026-08-10T02:44:31Z | 2026-04-25T20:49:00.733Z | 1 | Jackrong/Qwen3.5-9B-Claude-4.6-Opus-Reasoning-Distilled-GGUF | main | Qwen3.5-9B-Claude-4.6-Opus-Reasoning-Distilled-GGUF | Qwen/Qwen3.5-9B | Qwen | 9 | null | false | llama.cpp | b8639 | rocm | Q4_K_S | RX 7900 XTX | DISCRETE_GPU | RX 7900 XTX | 1 | 24 | null | null | null | null | null | 512 | 128 | 640 | 1 | null | null | 93.2 | null | null | null | null | null | null | [] | null | null | null | f16 | null | true | false | false | itgoatee | false | llama-bench on ROCm (hip-radeon); tg128 metric; b/ub 4096; f16/f16 KV cache |
external-community:cmoethtq9000pjz045yguc0ld | external-community | cmoethtq9000pjz045yguc0ld | 2026-08-10T02:44:31Z | 2026-04-25T20:54:13.185Z | 1 | Jackrong/Qwopus3.5-27B-v3-GGUF | main | Qwopus3.5-27B-v3-GGUF | unsloth/Qwen3.5-27B | Qwen | 27 | null | false | llama.cpp | b8778 | rocm | Q4_K_S | RX 7900 XTX | DISCRETE_GPU | RX 7900 XTX | 1 | 24 | null | null | null | null | null | 512 | 128 | 640 | 1 | null | null | 28.6 | null | null | null | null | null | 14.51 | [] | null | null | null | q4_0 | null | true | false | false | itgoatee | false | llama-bench on ROCm (hip-radeon); tg128 metric; b/ub 512; q4_0/q4_0 KV; gen perfectly flat tg128-tg2048 |
external-community:cmoetoim3000sjo04hd67abew | external-community | cmoetoim3000sjo04hd67abew | 2026-08-10T02:44:31Z | 2026-04-25T20:59:25.371Z | 1 | unsloth/Qwen3.6-27B-GGUF | main | Qwen3.6-27B-GGUF | Qwen/Qwen3.6-27B | Qwen | 27 | null | false | llama.cpp | b8778 | rocm | Q4_K_S | RX 7900 XTX | DISCRETE_GPU | RX 7900 XTX | 1 | 24 | null | null | null | null | null | 512 | 128 | 640 | 1 | null | null | 28.5 | null | null | null | null | null | 14.76 | [] | null | null | null | q4_0 | null | true | false | false | itgoatee | false | llama-bench on ROCm (hip-radeon); tg128 metric; b=1024/ub=512; q4_0/q4_0 KV; GatedDeltaNet+GQA hybrid arch; gen perfectly flat |
external-community:cmoetv7zn0004l704ocb4a949 | external-community | cmoetv7zn0004l704ocb4a949 | 2026-08-10T02:44:31Z | 2026-04-25T21:04:38.195Z | 1 | Jackrong/Gemopus-4-26B-A4B-it-GGUF | main | Gemopus-4-26B-A4B-it-GGUF | null | Llama | 26 | 4 | true | llama.cpp | b8778 | rocm | Q4_K_M | RX 7900 XTX | DISCRETE_GPU | RX 7900 XTX | 1 | 24 | null | null | null | null | null | 512 | 128 | 640 | 1 | null | null | 94.5 | null | null | null | null | null | 15.63 | [] | null | null | null | f16 | null | false | false | false | itgoatee | false | llama-bench on ROCm (hip-radeon); tg128 metric; b/ub 512; FA=0 required (Gemma-4 head_dim=512 exceeds FA2 limit; FA=1 degrades gen 50-70% on ROCm); f16/f16 KV mandatory with FA=0 |
external-community:cmoeu1wus0003l60472784yfr | external-community | cmoeu1wus0003l60472784yfr | 2026-08-10T02:44:31Z | 2026-04-25T21:09:50.356Z | 1 | unsloth/gemma-4-26B-A4B-it-GGUF | main | gemma-4-26B-A4B-it-GGUF | google/gemma-4-26B-A4B-it | Gemma | 26 | null | false | llama.cpp | b8778 | rocm | Q4_K_S | RX 7900 XTX | DISCRETE_GPU | RX 7900 XTX | 1 | 24 | null | null | null | null | null | 512 | 128 | 640 | 1 | null | null | 94.3 | null | null | null | null | null | 15.25 | [] | null | null | null | f16 | null | false | false | false | itgoatee | false | llama-bench on ROCm (hip-radeon); tg128 metric; b/ub 512; FA=0 required on ROCm; f16/f16 KV only (quantized V requires FA); UD unsloth dynamic quant |
external-community:cmoeu8m92000fl604nt2wtvkj | external-community | cmoeu8m92000fl604nt2wtvkj | 2026-08-10T02:44:31Z | 2026-04-25T21:15:03.206Z | 1 | lmstudio-community/Qwen3-Coder-30B-A3B-Instruct-GGUF | main | Qwen3-Coder-30B-A3B-Instruct-GGUF | Qwen/Qwen3-Coder-30B-A3B-Instruct | Qwen | 30 | null | false | llama.cpp | b8639 | rocm | Q4_K_M | RX 7900 XTX | DISCRETE_GPU | RX 7900 XTX | 1 | 24 | null | null | null | null | null | 512 | 512 | 1,024 | 1 | null | null | 46.9 | null | null | null | null | null | null | [] | null | null | null | q8_0 | null | true | false | false | itgoatee | false | llama-bench on ROCm (hip-radeon); tg512 metric; b/ub 1024; q8_0/q8_0 KV; MoE A3B (30B total / 3B active) |
external-community:cmoeufbnr000ql604wctafc0s | external-community | cmoeufbnr000ql604wctafc0s | 2026-08-10T02:44:31Z | 2026-04-25T21:20:16.072Z | 1 | cyberjuju/Qwen3-30B-A3B-Instruct-2507-AWQ-4bit-Q4_K_M-GGUF | main | Qwen3-30B-A3B-Instruct-2507-AWQ-4bit-Q4_K_M-GGUF | cyankiwi/Qwen3-30B-A3B-Instruct-2507-AWQ-4bit | Qwen | 30 | null | false | llama.cpp | b8639 | rocm | Q4_K_M | RX 7900 XTX | DISCRETE_GPU | RX 7900 XTX | 1 | 24 | null | null | null | null | null | 512 | 512 | 1,024 | 1 | null | null | 47.8 | null | null | null | null | null | null | [] | null | null | null | q4_0 | null | true | false | false | itgoatee | false | llama-bench on ROCm (hip-radeon); tg512 metric; b/ub 1024; q4_0/q4_0 KV; AWQ 4-bit re-quanted to Q4_K_M GGUF; MoE A3B |
external-community:cmoeum17e0003jo04iijt0vvl | external-community | cmoeum17e0003jo04iijt0vvl | 2026-08-10T02:44:31Z | 2026-04-25T21:25:29.115Z | 1 | DJLougen/Ornstein3.6-35B-A3B-GGUF | main | Ornstein3.6-35B-A3B-GGUF | DJLougen/Ornstein3.6-35B-A3B | Qwen | 35 | 3 | true | llama.cpp | b8778 | rocm | Q4_K_S | RX 7900 XTX | DISCRETE_GPU | RX 7900 XTX | 1 | 24 | null | null | null | null | null | 512 | 128 | 640 | 1 | null | null | 62.4 | null | null | null | null | null | null | [] | null | null | null | q8_0 | null | true | false | false | itgoatee | false | llama-bench on ROCm (hip-radeon); tg128 metric; b/ub 1024; q8_0/q8_0 KV; MoE A3B; gen perfectly flat (Qwen3.6 DeltaNet arch); Q4_K_S beats Q4_K_M on all metrics |
external-community:cmoeusq0f0007l804y0oofytc | external-community | cmoeusq0f0007l804y0oofytc | 2026-08-10T02:44:31Z | 2026-04-25T21:30:41.200Z | 1 | unsloth/Qwen3.6-35B-A3B-GGUF | main | Qwen3.6-35B-A3B-GGUF | Qwen/Qwen3.6-35B-A3B | Qwen | 35 | null | false | llama.cpp | b8778 | rocm | Q4_K_S | RX 7900 XTX | DISCRETE_GPU | RX 7900 XTX | 1 | 24 | null | null | null | null | null | 512 | 128 | 640 | 1 | null | null | 57.6 | null | null | null | null | null | 19.45 | [] | null | null | 33 | q4_0 | null | true | false | false | itgoatee | false | llama-bench on ROCm (hip-radeon); tg128 metric; b/ub 1024; q4_0/q4_0 KV; -ot cutoff=33 (layers 0-32 GPU / 33-39 CPU) gives 3x better prefill vs baseline; UD unsloth dynamic quant; MoE A3B |
external-community:cmoevcpkp000sl804pb3yoj0c | external-community | cmoevcpkp000sl804pb3yoj0c | 2026-08-10T02:44:31Z | 2026-04-25T21:46:13.753Z | 1 | sakamakismile/Qwen3.6-27B-NVFP4 | main | Qwen3.6-27B-NVFP4 | Qwen/Qwen3.6-27B | Qwen | 27 | null | false | vllm | 0.19.2rc1.dev134+gfe9c3d6c5 | cuda | NVFP4 | RTX 5090 | DISCRETE_GPU | RTX 5090 | 1 | 32 | null | null | Ryzen 7 3700X | null | Ubuntu 24.04 | 128 | 128 | 131,072 | 4 | null | 258.05 | 57.54 | null | 57.54 | null | null | null | 28.1 | [] | null | null | null | fp8 | null | null | false | false | rohit | false | vLLM on Unraid RTX 5090 (600W). NVFP4 quant, fp8 KV cache, prefix caching, 131K context. Updated after power limit increase — ~13% faster output, ~17% faster TTFT. |
external-community:cmoevgkv6000iib04cmisudtv | external-community | cmoevgkv6000iib04cmisudtv | 2026-08-10T02:44:31Z | 2026-04-25T21:49:14.274Z | 1 | unsloth/Qwen3.6-27B-GGUF | main | Qwen3.6-27B-GGUF | Qwen/Qwen3.6-27B | Qwen | 27 | null | false | llama.cpp | null | null | Q4_K_M | RTX 5090 | DISCRETE_GPU | RTX 5090 | 1 | 32 | null | null | Ryzen 7 3700X | null | Ubuntu 24.04 | 0 | 0 | 2,048 | 1 | null | 459 | 89 | null | null | null | null | null | null | [] | null | null | null | null | null | null | false | false | claudehath | false | null |
external-community:cmoevn4cu0016l804vaasia9h | external-community | cmoevn4cu0016l804vaasia9h | 2026-08-10T02:44:31Z | 2026-04-25T21:54:19.471Z | 1 | unsloth/Qwen3.6-27B-GGUF | main | Qwen3.6-27B-GGUF | Qwen/Qwen3.6-27B | Qwen | 27 | null | false | llama.cpp | null | null | Q4_K_M | RTX 5090 | DISCRETE_GPU | RTX 5090 | 1 | 32 | null | null | Ryzen 7 3700X | null | Ubuntu 24.04 | 0 | 0 | 262,144 | 1 | null | 459 | 89 | null | null | null | null | null | null | [] | null | null | 99 | q8_0 | null | true | false | false | claudehath | false | null |
external-community:cmof1zrjo0003h204re6e99ao | external-community | cmof1zrjo0003h204re6e99ao | 2026-08-10T02:44:31Z | 2026-04-26T00:52:07.092Z | 1 | Qwen/Qwen3.6-35B-A3B | main | Qwen3.6-35B-A3B | null | Qwen | 36 | 3 | true | vllm | 0.19.2rc1.dev21+g893611813 | cuda | INT4 | RTX 3090 | DISCRETE_GPU | RTX 3090 | 2 | 24 | null | null | AMD EPYC 7543 | null | Ubuntu 24.04 | 0 | 0 | 32,768 | 1 | null | null | 170.5 | null | null | null | null | null | 22.5 | [] | null | 2 | null | fp16 | null | null | false | false | wasifb | false | Quant: Intel/Qwen3.6-35B-A3B-int4-AutoRound (group_size=128, sym, W4A16, repo currently gated on HF — used while it was public). Requires our pad-Marlin patch (vLLM PR #40361, OPEN) for TP=2 — without it Marlin rejects size_n=32 sub-tile shards. Measured 2026-04-20 at 230W power cap on consumer RTX 3090s, no NVLink bri... |
external-community:cmof29ivq000ajr047dp6behr | external-community | cmof29ivq000ajr047dp6behr | 2026-08-10T02:44:31Z | 2026-04-26T00:59:42.423Z | 1 | unsloth/Qwen3.6-35B-A3B-GGUF | main | Qwen3.6-35B-A3B-GGUF | Qwen/Qwen3.6-35B-A3B | Qwen | 35 | null | false | llama.cpp | turboquant-fork | cuda | UD-Q4_K_XL | RTX 3090 | DISCRETE_GPU | RTX 3090 | 2 | 24 | null | null | AMD EPYC 7543 | null | Ubuntu 24.04 | 0 | 0 | 1,048,576 | 1 | null | null | 124 | null | null | null | null | null | 22.5 | [] | null | null | 999 | q4_0 | null | true | false | false | wasifb | false | Qwen3.6-35B-A3B (MoE, 3B-active) at 1M context window. Layer-split across 2× RTX 3090, llama.cpp turboquant fork (Tom Jobbins' fork supporting q4_0 KV cache + extended ctx). Short-prompt 124 TPS sustained at full 1M ctx. Loaded-context generation 37 TPS @ 128K. UD-Q4_K_XL = Unsloth Dynamic Quantization (variable-precis... |
external-community:cmof7f8y90002ju04kejdop3a | external-community | cmof7f8y90002ju04kejdop3a | 2026-08-10T02:44:31Z | 2026-04-26T03:24:07.569Z | 1 | unsloth/Qwen3.5-35B-A3B-GGUF | main | Qwen3.5-35B-A3B-GGUF | Qwen/Qwen3.5-35B-A3B | Qwen | 35 | null | false | llama.cpp | null | null | UD-Q4_K_XL | RTX 4090 | DISCRETE_GPU | RTX 4090 | 1 | 24 | null | null | Intel Core Ultra 9 285K | null | Fedora 43 | 0 | 0 | 2,048 | 1 | null | null | 157.4 | null | null | null | null | null | null | [] | null | null | null | null | null | null | false | false | basecampbernie | false | null |
external-community:cmof7ls5a0003l604jnromcue | external-community | cmof7ls5a0003l604jnromcue | 2026-08-10T02:44:31Z | 2026-04-26T03:29:12.382Z | 1 | unsloth/Qwen3.5-35B-A3B-GGUF | main | Qwen3.5-35B-A3B-GGUF | Qwen/Qwen3.5-35B-A3B | Qwen | 35 | null | false | llama.cpp | null | null | UD-Q4_K_XL | RTX 3090 | DISCRETE_GPU | RTX 3090 | 1 | 24 | null | null | AMD EPYC 7543 | null | Ubuntu 24.04 | 0 | 0 | 2,048 | 1 | null | null | 124.4 | null | null | null | null | null | null | [] | null | null | null | null | null | null | false | false | basecampbernie | false | BigBeast single RTX 3090, Qwen3.5-35B-A3B MoE, Q8_0 KV, single slot. Prompt processing 811 tok/s. Source: ~/aurora_docs/QWEN35_LOCAL_BENCHMARKS.md on bigbeast-fedora. |
external-community:cmof7sd4b0002l40473uvzk0u | external-community | cmof7sd4b0002l40473uvzk0u | 2026-08-10T02:44:31Z | 2026-04-26T03:34:19.499Z | 1 | unsloth/Qwen3.5-27B-GGUF | main | Qwen3.5-27B-GGUF | Qwen/Qwen3.5-27B | Qwen | 27 | null | false | llama.cpp | null | null | Q4_0 | RTX 4090 | DISCRETE_GPU | RTX 4090 | 1 | 24 | null | null | Intel Core Ultra 9 285K | null | Fedora 43 | 0 | 0 | 2,048 | 1 | null | null | 44 | null | null | null | null | null | null | [] | null | null | null | null | null | null | false | false | basecampbernie | false | BigBeast single RTX 4090, Qwen3.5-27B dense Q4_0, Q4_1 KV due 256K VRAM fit. Prompt processing 1105 tok/s. Source: ~/aurora_docs/QWEN35_LOCAL_BENCHMARKS.md. |
external-community:cmof85if50007l4048uhps6jx | external-community | cmof85if50007l4048uhps6jx | 2026-08-10T02:44:31Z | 2026-04-26T03:44:32.897Z | 1 | unsloth/Qwen3.5-27B-GGUF | main | Qwen3.5-27B-GGUF | Qwen/Qwen3.5-27B | Qwen | 27 | null | false | llama.cpp | null | null | Q4_K_M | RTX 4090 | DISCRETE_GPU | RTX 4090 | 1 | 24 | null | null | Intel Core Ultra 9 285K | null | Fedora 43 | 0 | 0 | 2,048 | 1 | null | null | 46.69 | null | null | null | null | null | null | [] | null | null | null | null | null | null | false | false | basecampbernie | false | BigBeast TurboQuant turbo3 KV, Qwen3.5-27B Q4_K_M on RTX 4090. pp512 3021 tok/s, tg128 46.69 tok/s. Source: ~/aurora_docs/BIGBEAST_TURBO3_BENCHMARKS.md. |
external-community:cmof9h1970008l704j9dc7oo7 | external-community | cmof9h1970008l704j9dc7oo7 | 2026-08-10T02:44:31Z | 2026-04-26T04:21:30.140Z | 1 | Qwen/Qwen3.6-35B-A3B-FP8 | main | Qwen3.6-35B-A3B-FP8 | Qwen/Qwen3.6-35B-A3B | Qwen | 35 | null | false | vllm | 0.19.1 | cuda | fp8 | RTX 3090 | DISCRETE_GPU | RTX 3090 | 2 | 24 | null | null | AMD EPYC 7543 | null | Ubuntu 24.04 | 512 | 512 | 32,768 | 1 | null | 148.56 | 140.01 | null | 280.03 | null | null | null | 22.15 | [] | null | 2 | null | fp8 | null | null | false | false | Skiipy | false | Qwen 3.6 35B-A3B MoE (3B active per token, sparse activation). Mixed-pair rig: RTX 3090 Ti (slot 1, x16) + RTX 3090 FE (slot 2, x4); TP=2 still gated by FE bandwidth. Single-stream 512/512 via `vllm bench serve` at concurrency=1, 20 prompts. Median TPOT 6.86 ms (~146 t/s decode). Required `--max-num-batched-tokens 4096... |
external-community:cmofapunx0008l20407bq5ln9 | external-community | cmofapunx0008l20407bq5ln9 | 2026-08-10T02:44:31Z | 2026-04-26T04:56:21.117Z | 1 | Qwen/Qwen3.6-35B-A3B | main | Qwen3.6-35B-A3B | null | Qwen | 36 | 3 | true | llama.cpp | b8965 | cuda | UD-Q4_K_XL | RTX 5090 | DISCRETE_GPU | RTX 5090 | 1 | 32 | null | null | Ryzen 7 3700X | null | Ubuntu 24.04 | 34 | 1,499 | 262,144 | 1 | null | 54 | 190 | null | null | null | null | null | 30 | [] | null | null | null | null | null | null | false | false | claudehath | false | TurboQuant2 fork (TheTom/llama-cpp-turboquant). cache-type-k q8_0, cache-type-v q8_0, flash-attn on, ctx 262144, thinking enabled (Qwen3 reasoning mode), mlock, no-mmap, prio 2 |
external-community:cmofbm4i90004jt04iovaovs7 | external-community | cmofbm4i90004jt04iovaovs7 | 2026-08-10T02:44:31Z | 2026-04-26T05:21:26.865Z | 1 | unsloth/Qwen3.5-9B-GGUF | main | Qwen3.5-9B-GGUF | Qwen/Qwen3.5-9B | Qwen | 9 | null | false | llama.cpp | b8934 | null | Q4_K_S | RTX 3090 | DISCRETE_GPU | RTX 3090 | 2 | 24 | null | null | AMD EPYC 7543 | null | Ubuntu 24.04 | 512 | 128 | 8,192 | 1 | null | null | 110.11 | null | 3,842.59 | null | null | null | null | [] | null | null | null | null | null | null | false | false | null | false | 2x RTX 3090 NVLink, all layers on GPU, flash attention. llama-bench pp512 tg128. |
external-community:cmofbmhhd0007jt04k1tmw4hb | external-community | cmofbmhhd0007jt04k1tmw4hb | 2026-08-10T02:44:31Z | 2026-04-26T05:21:43.681Z | 1 | unsloth/Qwen3.6-27B-GGUF | main | Qwen3.6-27B-GGUF | Qwen/Qwen3.6-27B | Qwen | 27 | null | false | llama.cpp | null | null | UD-Q4_K_XL | RTX 4090 | DISCRETE_GPU | RTX 4090 | 1 | 24 | null | null | Intel Core Ultra 9 285K | null | Fedora 43 | 2,048 | 128 | 32,768 | 1 | null | null | 41.82 | null | null | null | null | null | null | [] | null | null | null | null | null | null | false | false | basecampbernie | false | BigBeast RTX 4090 Qwen3.6-27B UD-Q4_K_XL GGUF, q4_1/q4_1 KV, flash attention, llama.cpp build 8892. Batched bench B=1: pp2048 2981 tok/s, tg128 41.82 tok/s. Source: /home/radar/llama.cpp/logs/bench/qwen36-27b-q4-batched-4090-20260422_215014.md on bigbeast-fedora. |
external-community:cmofbxetu0009l504eulxcb9j | external-community | cmofbxetu0009l504eulxcb9j | 2026-08-10T02:44:31Z | 2026-04-26T05:30:13.458Z | 1 | dealignai/Qwen3.6-27B-MXFP4-CRACK | main | Qwen3.6-27B-MXFP4-CRACK | Qwen/Qwen3.6-27B | Qwen | 27 | null | false | mlx | null | null | MXFP4 | M4 Max | UNIFIED | null | 1 | null | 128 | null | null | null | null | 0 | 0 | 2,048 | 1 | null | null | 33 | null | null | null | null | null | null | [] | null | null | null | null | null | null | false | false | hornsan1 | false | null |
external-community:cmofcsjz8000fjo046uao6hnx | external-community | cmofcsjz8000fjo046uao6hnx | 2026-08-10T02:44:31Z | 2026-04-26T05:54:26.468Z | 1 | GestaltLabs/Ornstein-3.6-27B-GGUF | main | Ornstein-3.6-27B-GGUF | GestaltLabs/Ornstein-3.6-27B | Qwen | 27 | null | false | llama.cpp | null | null | Q4_K_M | M4 Max | UNIFIED | null | 1 | null | 128 | null | null | null | null | 0 | 0 | 32,768 | 1 | null | null | 22 | null | null | null | null | null | null | [] | null | null | null | null | null | null | false | false | hornsan1 | false | llama-server Metal, --flash-attn on, --n-gpu-layers 999, thinking enabled |
external-community:cmofcz3qg000pl7041jiw94ba | external-community | cmofcz3qg000pl7041jiw94ba | 2026-08-10T02:44:31Z | 2026-04-26T05:59:32.008Z | 1 | mlx-community/gemma-4-26b-a4b-it-4bit | main | gemma-4-26b-a4b-it-4bit | google/gemma-4-26B-A4B-it | Gemma | 26 | null | false | mlx | null | null | 4bit | M4 Max | UNIFIED | null | 1 | null | 128 | null | null | null | null | 0 | 256 | 2,048 | 1 | null | null | 117 | null | null | null | null | null | null | [] | null | null | null | null | null | null | false | false | hornsan1 | false | mlx_vlm.server, 256-tok generation. peak generation_tps 117 (4-tok), sustained ~102 over 256-tok runs. MoE 4B-active. |
external-community:cmofd5ok6000ql5046j1r55pn | external-community | cmofd5ok6000ql5046j1r55pn | 2026-08-10T02:44:31Z | 2026-04-26T06:04:38.934Z | 1 | mlx-community/Qwen3.6-35B-A3B-4bit | main | Qwen3.6-35B-A3B-4bit | Qwen/Qwen3.6-35B-A3B | Qwen | 6 | null | false | mlx | null | null | 4bit | M4 Max | UNIFIED | null | 1 | null | 128 | null | null | null | null | 0 | 16 | 2,048 | 1 | null | null | 109 | null | null | null | null | null | null | [] | null | null | null | null | null | null | false | false | hornsan1 | false | mlx_vlm.server, post-warmup short-burst peak. 256-tok sustained ~97 tok/s. MoE 3B-active / 35B total. Peak memory 20.5GB. |
external-community:cmofdidib0002ju04tsntnze1 | external-community | cmofdidib0002ju04tsntnze1 | 2026-08-10T02:44:31Z | 2026-04-26T06:14:31.139Z | 1 | Qwen/Qwen3.5-9B | main | Qwen3.5-9B | Qwen/Qwen3.5-9B-Base | Qwen | 9 | null | false | hipfire | 0.1.8-alpha+f16eceb | rocm | MQ4 | RX 7900 XTX | DISCRETE_GPU | RX 7900 XTX | 1 | 24 | null | null | null | null | null | 20 | 128 | 4,096 | 1 | null | 48.1 | 122.3 | null | 116.9 | null | null | null | 5.662109 | [] | null | null | null | null | null | null | false | false | schuttdev | false | hipfire @ f16eceb (master post-perf-recovery)
prompt: "Explain the theory of general relativity in simple terms."
runs: 3
decode mode: AR
kv_cache: asym3
prompt_normalize: true (default since 2026-04-26)
pp128: 602.8 tok/s
pp512: 591.4 tok/s
pp1024: 587.8 tok/s
pp2048: 580.0 tok/s
vram_loaded: 5798 MB |
external-community:cmofdju0x0006ju041pympxzf | external-community | cmofdju0x0006ju041pympxzf | 2026-08-10T02:44:31Z | 2026-04-26T06:15:39.202Z | 1 | DJLougen/Ornstein3.6-35B-A3B-GGUF | main | Ornstein3.6-35B-A3B-GGUF | DJLougen/Ornstein3.6-35B-A3B | Qwen | 35 | 3 | true | llama.cpp | b8934 | null | Q4_K_M | RTX 3090 | DISCRETE_GPU | RTX 3090 | 2 | 24 | null | null | AMD EPYC 7543 | null | Ubuntu 24.04 | 512 | 128 | 8,192 | 1 | null | null | 138.01 | null | 3,335.33 | null | null | null | null | [] | null | null | null | null | null | null | false | false | null | false | 2x RTX 3090 NVLink, Ornstein 3.6-35B-A3B MoE (3B active params), all layers on GPU, flash attention. llama-bench pp512 tg128. |
external-community:cmofdp1ve0003i904tw2yf6bs | external-community | cmofdp1ve0003i904tw2yf6bs | 2026-08-10T02:44:31Z | 2026-04-26T06:19:42.650Z | 1 | Qwen/Qwen3.5-9B | main | Qwen3.5-9B | Qwen/Qwen3.5-9B-Base | Qwen | 9 | null | false | hipfire | 0.1.8-alpha+f16eceb | rocm | MQ4 | RX 7900 XTX | DISCRETE_GPU | RX 7900 XTX | 1 | 24 | null | null | null | null | null | 220 | 120 | 4,096 | 1 | null | null | 336.98 | null | null | null | null | null | null | [] | null | null | null | null | null | null | true | false | schuttdev | false | hipfire @ f16eceb (master post-perf-recovery, PR #47)
prompt: LRU cache PEP-8 strict (md5=df5dedc8040c, see benchmarks/prompts/lru_cache_pep8_strict.txt)
runs: 3 (median reported)
decode mode: DFlash speculative (block_size=16)
kv_cache: asym3
prompt_normalize: true (default since 2026-04-26)
τ (median): 7.5
accept_rat... |
external-community:cmofdvq7c0001l40461v8g3z7 | external-community | cmofdvq7c0001l40461v8g3z7 | 2026-08-10T02:44:31Z | 2026-04-26T06:24:54.120Z | 1 | Qwen/Qwen3.5-27B | main | Qwen3.5-27B | null | Qwen | 28 | null | false | hipfire | 0.1.8-alpha+f16eceb | rocm | MQ4 | RX 7900 XTX | DISCRETE_GPU | RX 7900 XTX | 1 | 24 | null | null | null | null | null | 20 | 128 | 4,096 | 1 | null | 117.8 | 43.7 | null | 42 | null | null | null | 15.105469 | [] | null | null | null | null | null | null | false | false | schuttdev | false | hipfire @ f16eceb (master post-perf-recovery)
prompt: "Explain the theory of general relativity in simple terms."
runs: 3
decode mode: AR
kv_cache: asym3
prompt_normalize: true (default since 2026-04-26)
pp128: 223.5 tok/s
pp512: 215.7 tok/s
pp1024: 213.3 tok/s
pp2048: 210.9 tok/s
vram_loaded: 15468 MB |
external-community:cmofe2efg000kla04k5h7rto0 | external-community | cmofe2efg000kla04k5h7rto0 | 2026-08-10T02:44:31Z | 2026-04-26T06:30:05.453Z | 1 | Qwen/Qwen3.5-27B | main | Qwen3.5-27B | null | Qwen | 28 | null | false | hipfire | 0.1.8-alpha+f16eceb | rocm | MQ4 | RX 7900 XTX | DISCRETE_GPU | RX 7900 XTX | 1 | 24 | null | null | null | null | null | 220 | 120 | 4,096 | 1 | null | null | 181.96 | null | null | null | null | null | null | [] | null | null | null | null | null | null | true | false | schuttdev | false | hipfire @ f16eceb (master post-perf-recovery, PR #47)
prompt: LRU cache PEP-8 strict (md5=df5dedc8040c, see benchmarks/prompts/lru_cache_pep8_strict.txt)
runs: 3 (median reported)
decode mode: DFlash speculative (block_size=16)
kv_cache: asym3
prompt_normalize: true (default since 2026-04-26)
τ (median): 9.417
accept_r... |
external-community:cmofe5y0g0003l704etimxv7c | external-community | cmofe5y0g0003l704etimxv7c | 2026-08-10T02:44:31Z | 2026-04-26T06:32:50.801Z | 1 | unsloth/Qwen3.6-35B-A3B-GGUF | main | Qwen3.6-35B-A3B-GGUF | Qwen/Qwen3.6-35B-A3B | Qwen | 35 | null | false | llama.cpp | build_info: b8925-0adede866 | cuda | Q4_K_M | RTX 4090 | DISCRETE_GPU | RTX 4090 | 1 | 24 | null | null | i9 14900k | null | Ubuntu 24.04 | 512 | 512 | 64,000 | 1 | null | null | 165.06 | null | null | null | null | null | 23,691 | [] | null | null | null | null | null | null | false | false | adeebaldkheel | false | null |
external-community:cmofe92tt000pla04ivnjv4jd | external-community | cmofe92tt000pla04ivnjv4jd | 2026-08-10T02:44:31Z | 2026-04-26T06:35:17.010Z | 1 | Qwen/Qwen3.5-35B-A3B | main | Qwen3.5-35B-A3B | Qwen/Qwen3.5-35B-A3B-Base | Qwen | 35 | null | false | hipfire | 0.1.8-alpha+f16eceb | rocm | MQ4 | RX 7900 XTX | DISCRETE_GPU | RX 7900 XTX | 1 | 24 | null | null | null | null | null | 20 | 128 | 4,096 | 1 | null | 75.4 | 135.9 | null | 125.9 | null | null | null | 22.128906 | [] | null | null | null | null | null | null | false | false | schuttdev | false | hipfire @ f16eceb (master post-perf-recovery)
prompt: "Explain the theory of general relativity in simple terms."
runs: 3
decode mode: AR
kv_cache: asym3
prompt_normalize: true (default since 2026-04-26)
pp128: 326.1 tok/s
pp512: 322.5 tok/s
pp1024: 320.8 tok/s
pp2048: 317.5 tok/s
vram_loaded: 22660 MB |
external-community:cmofefr0j000ula04chaak8ou | external-community | cmofefr0j000ula04chaak8ou | 2026-08-10T02:44:31Z | 2026-04-26T06:40:28.291Z | 1 | Qwen/Qwen3.5-35B-A3B | main | Qwen3.5-35B-A3B | Qwen/Qwen3.5-35B-A3B-Base | Qwen | 35 | null | false | hipfire | 0.1.8-alpha+f16eceb | rocm | MQ4 | RX 7900 XTX | DISCRETE_GPU | RX 7900 XTX | 1 | 24 | null | null | null | null | null | 220 | 120 | 4,096 | 1 | null | null | 140.59 | null | null | null | null | null | null | [] | null | null | null | null | null | null | true | false | schuttdev | false | hipfire @ f16eceb (master post-perf-recovery, PR #47)
prompt: LRU cache PEP-8 strict (md5=df5dedc8040c, see benchmarks/prompts/lru_cache_pep8_strict.txt)
runs: 3 (median reported)
decode mode: DFlash speculative (block_size=16)
kv_cache: asym3
prompt_normalize: true (default since 2026-04-26)
τ (median): 4.905
accept_r... |
external-community:cmofemf840014la04lhmues5d | external-community | cmofemf840014la04lhmues5d | 2026-08-10T02:44:31Z | 2026-04-26T06:45:39.604Z | 1 | Qwen/Qwen3.6-27B | main | Qwen3.6-27B | null | Qwen | 28 | null | false | hipfire | 0.1.8-alpha+f16eceb | rocm | MQ4 | RX 7900 XTX | DISCRETE_GPU | RX 7900 XTX | 1 | 24 | null | null | null | null | null | 20 | 128 | 4,096 | 1 | null | 117.6 | 43.6 | null | 41.9 | null | null | null | 15.105469 | [] | null | null | null | null | null | null | false | false | schuttdev | false | hipfire @ f16eceb (master post-perf-recovery)
prompt: "Explain the theory of general relativity in simple terms."
runs: 3
decode mode: AR
kv_cache: asym3
prompt_normalize: true (default since 2026-04-26)
pp128: 224.2 tok/s
pp512: 216.5 tok/s
pp1024: 214.3 tok/s
pp2048: 211.1 tok/s
vram_loaded: 15468 MB |
external-community:cmofet3j80005l2042r92gywi | external-community | cmofet3j80005l2042r92gywi | 2026-08-10T02:44:31Z | 2026-04-26T06:50:51.045Z | 1 | Qwen/Qwen3.6-27B | main | Qwen3.6-27B | null | Qwen | 28 | null | false | hipfire | 0.1.8-alpha+f16eceb | rocm | MQ4 | RX 7900 XTX | DISCRETE_GPU | RX 7900 XTX | 1 | 24 | null | null | null | null | null | 220 | 120 | 4,096 | 1 | null | null | 118.16 | null | null | null | null | null | null | [] | null | null | null | null | null | null | true | false | schuttdev | false | hipfire @ f16eceb (master post-perf-recovery, PR #47)
prompt: LRU cache PEP-8 strict (md5=df5dedc8040c, see benchmarks/prompts/lru_cache_pep8_strict.txt)
runs: 3 (median reported)
decode mode: DFlash speculative (block_size=16)
kv_cache: asym3
prompt_normalize: true (default since 2026-04-26)
τ (median): 5.632
accept_r... |
external-community:cmofezrun000cie04z5y0b83e | external-community | cmofezrun000cie04z5y0b83e | 2026-08-10T02:44:31Z | 2026-04-26T06:56:02.495Z | 1 | Qwen/Qwen3.6-35B-A3B | main | Qwen3.6-35B-A3B | null | Qwen | 36 | 3 | true | hipfire | 0.1.8-alpha+f16eceb | rocm | MQ4 | RX 7900 XTX | DISCRETE_GPU | RX 7900 XTX | 1 | 24 | null | null | null | null | null | 20 | 128 | 4,096 | 1 | null | 75.5 | 135.3 | null | 125.3 | null | null | null | 22.128906 | [] | null | null | null | null | null | null | false | false | schuttdev | false | hipfire @ f16eceb (master post-perf-recovery)
prompt: "Explain the theory of general relativity in simple terms."
runs: 3
decode mode: AR
kv_cache: asym3
prompt_normalize: true (default since 2026-04-26)
pp128: 325.4 tok/s
pp512: 322.0 tok/s
pp1024: 320.4 tok/s
pp2048: 316.8 tok/s
vram_loaded: 22660 MB |
external-community:cmoff6g560019la04hewfjrll | external-community | cmoff6g560019la04hewfjrll | 2026-08-10T02:44:31Z | 2026-04-26T07:01:13.915Z | 1 | Qwen/Qwen3.6-35B-A3B | main | Qwen3.6-35B-A3B | null | Qwen | 36 | 3 | true | hipfire | 0.1.8-alpha+f16eceb | rocm | MQ4 | RX 7900 XTX | DISCRETE_GPU | RX 7900 XTX | 1 | 24 | null | null | null | null | null | 220 | 120 | 4,096 | 1 | null | null | 68.6 | null | null | null | null | null | null | [] | null | null | null | null | null | null | true | false | schuttdev | false | hipfire @ f16eceb (master post-perf-recovery, PR #47)
prompt: LRU cache PEP-8 strict (md5=df5dedc8040c, see benchmarks/prompts/lru_cache_pep8_strict.txt)
runs: 3 (median reported)
decode mode: DFlash speculative (block_size=16)
kv_cache: asym3
prompt_normalize: true (default since 2026-04-26)
τ (median): 1.222
accept_r... |
external-community:cmofgd9iz000ml4055xsdgvuy | external-community | cmofgd9iz000ml4055xsdgvuy | 2026-08-10T02:44:31Z | 2026-04-26T07:34:31.547Z | 1 | DJLougen/Ornstein3.6-35B-A3B | main | Ornstein3.6-35B-A3B | unsloth/Qwen3.6-35B-A3B | Qwen | 35 | 3 | true | hipfire | 0.1.8-alpha+f16eceb | rocm | MQ4 | RX 7900 XTX | DISCRETE_GPU | RX 7900 XTX | 1 | 24 | null | null | null | null | null | 20 | 128 | 4,096 | 1 | null | 75.6 | 134 | null | 124.2 | null | null | null | 22.13 | [] | null | null | null | null | null | null | false | false | schuttdev | false | hipfire @ 0.1.8-alpha+f16eceb (master post-perf-recovery, PR #47)
prompt: "Explain the theory of general relativity in simple terms."
runs: 3 (median reported)
decode mode: AR (autoregressive, no spec-decode)
kv_cache: asym3 (3-bit rotated K + Q8 V; 5.5x vs fp32)
prompt_normalize: true (default since 2026-04-26)
pp128:... |
external-community:cmofi32rs000rjx04be81ns5r | external-community | cmofi32rs000rjx04be81ns5r | 2026-08-10T02:44:31Z | 2026-04-26T08:22:35.465Z | 1 | unsloth/Qwen3.6-27B-UD-MLX-4bit | main | Qwen3.6-27B-UD-MLX-4bit | Qwen/Qwen3.6-27B | Qwen | 27 | null | false | mlx | mlx-lm 0.31.3 / mlx 0.31.2 / mlx-metal 0.31.2 | metal | int4 | M4 Max | UNIFIED | null | 1 | null | 64 | null | Apple M4 Max 16-core CPU / 40-core GPU | null | macOS 26.4 build 25E246 | 128 | 6,000 | 6,128 | 1 | null | 676 | 18.435 | null | 18.786 | null | null | null | 25.895 | [] | null | null | null | null | null | null | false | false | shikhar | false | MLX benchmark Trial 1. TTFT is estimated from prompt_tps and generation_tps, not directly reported by mlx_lm.benchmark. Model config uses 4-bit affine group size 64 with selected 8-bit tensors. |
external-community:cmofia1vo0009l7048w39brkk | external-community | cmofia1vo0009l7048w39brkk | 2026-08-10T02:44:31Z | 2026-04-26T08:28:00.900Z | 1 | unsloth/Qwen3.6-27B-UD-MLX-4bit | main | Qwen3.6-27B-UD-MLX-4bit | Qwen/Qwen3.6-27B | Qwen | 27 | null | false | mlx | mlx-lm 0.31.3 / mlx 0.31.2 / mlx-metal 0.31.2 | metal | int4 | M4 Max | UNIFIED | null | 1 | null | 64 | null | Apple M4 Max 16-core CPU / 40-core GPU | null | macOS 26.4 build 25E246 | 128 | 6,000 | 6,128 | 1 | null | 676 | 18.434 | null | 18.786 | null | null | null | 25.896 | [] | null | null | null | null | null | null | false | false | shikhar | false | MLX benchmark 5-trial average. Trials: generation_tps 18.435, 18.433, 18.437, 18.421, 18.445; prompt_tps avg 205.970; generation_tps avg 18.434; peak_memory avg 25.896 GB. TTFT is estimated from avg prompt_tps and generation_tps because mlx_lm.benchmark does not directly report TTFT. Model config uses 4-bit affine grou... |
external-community:cmofkl1gm0006l4045txzjtoc | external-community | cmofkl1gm0006l4045txzjtoc | 2026-08-10T02:44:31Z | 2026-04-26T09:32:32.806Z | 1 | Qwen/Qwen3.5-9B | main | Qwen3.5-9B | Qwen/Qwen3.5-9B-Base | Qwen | 9 | null | false | hipfire | 0.1.8-alpha+f16eceb | rocm | MQ4 | RX 7900 XTX | DISCRETE_GPU | RX 7900 XTX | 1 | 24 | null | null | null | null | null | 220 | 120 | 4,096 | 1 | null | null | 322.6 | null | null | null | null | null | null | [] | null | null | null | null | null | null | true | false | schuttdev | false | hipfire @ f16eceb (master post-perf-recovery PR #47)
prompt: LRU cache PEP-8 strict (md5=df5dedc8040c, see benchmarks/prompts/lru_cache_pep8_strict.txt)
runs: 3 (median reported); range 322.5–346.0
decode mode: DFlash speculative (block_size=16)
kv_cache: asym3
prompt_normalize: true (default since 2026-04-26)
τ (media... |
external-community:cmofkrpp0000hic04ymyzumfg | external-community | cmofkrpp0000hic04ymyzumfg | 2026-08-10T02:44:31Z | 2026-04-26T09:37:44.148Z | 1 | Qwen/Qwen3.5-27B | main | Qwen3.5-27B | null | Qwen | 28 | null | false | hipfire | 0.1.8-alpha+f16eceb | rocm | MQ4 | RX 7900 XTX | DISCRETE_GPU | RX 7900 XTX | 1 | 24 | null | null | null | null | null | 220 | 120 | 4,096 | 1 | null | null | 201.11 | null | null | null | null | null | null | [] | null | null | null | null | null | null | true | false | schuttdev | false | hipfire @ f16eceb (master post-perf-recovery PR #47)
prompt: LRU cache PEP-8 strict (md5=df5dedc8040c, see benchmarks/prompts/lru_cache_pep8_strict.txt)
runs: 3 (median reported); range 201.0–201.1
decode mode: DFlash speculative (block_size=16)
kv_cache: asym3
prompt_normalize: true (default since 2026-04-26)
τ (media... |
external-community:cmofkye0g0009l404ap3uadyl | external-community | cmofkye0g0009l404ap3uadyl | 2026-08-10T02:44:31Z | 2026-04-26T09:42:55.600Z | 1 | Qwen/Qwen3.5-35B-A3B | main | Qwen3.5-35B-A3B | Qwen/Qwen3.5-35B-A3B-Base | Qwen | 35 | null | false | hipfire | 0.1.8-alpha+f16eceb | rocm | MQ4 | RX 7900 XTX | DISCRETE_GPU | RX 7900 XTX | 1 | 24 | null | null | null | null | null | 220 | 120 | 4,096 | 1 | null | null | 154.62 | null | null | null | null | null | null | [] | null | null | null | null | null | null | true | false | schuttdev | false | hipfire @ f16eceb (master post-perf-recovery PR #47)
prompt: LRU cache PEP-8 strict (md5=df5dedc8040c, see benchmarks/prompts/lru_cache_pep8_strict.txt)
runs: 3 (median reported); range 66.4–155.9
decode mode: DFlash speculative (block_size=16)
kv_cache: asym3
prompt_normalize: true (default since 2026-04-26)
τ (median... |
external-community:cmofl528t000nl4045c6nx68j | external-community | cmofl528t000nl4045c6nx68j | 2026-08-10T02:44:31Z | 2026-04-26T09:48:06.941Z | 1 | Qwen/Qwen3.6-27B | main | Qwen3.6-27B | null | Qwen | 28 | null | false | hipfire | 0.1.8-alpha+f16eceb | rocm | MQ4 | RX 7900 XTX | DISCRETE_GPU | RX 7900 XTX | 1 | 24 | null | null | null | null | null | 220 | 120 | 4,096 | 1 | null | null | 118.14 | null | null | null | null | null | null | [] | null | null | null | null | null | null | true | false | schuttdev | false | hipfire @ f16eceb (master post-perf-recovery PR #47)
prompt: LRU cache PEP-8 strict (md5=df5dedc8040c, see benchmarks/prompts/lru_cache_pep8_strict.txt)
runs: 3 (median reported); range 117.5–118.4
decode mode: DFlash speculative (block_size=16)
kv_cache: asym3
prompt_normalize: true (default since 2026-04-26)
τ (media... |
external-community:cmofps6zq0010l7040imyvvtx | external-community | cmofps6zq0010l7040imyvvtx | 2026-08-10T02:44:31Z | 2026-04-26T11:58:04.647Z | 1 | sakamakismile/Huihui-Qwen3.6-27B-abliterated-NVFP4-MTP | main | Huihui-Qwen3.6-27B-abliterated-NVFP4-MTP | huihui-ai/Huihui-Qwen3.6-27B-abliterated | Qwen | 27 | null | false | vllm | v0.19.2rc1.dev134+gfe9c3d6c5 | null | NVFP4 | RTX 5090 | DISCRETE_GPU | RTX 5090 | 1 | 32 | null | null | Ryzen 7 3700X | null | Ubuntu 24.04 | 21 | 1,000 | 131,072 | 1 | null | 108 | 97.1 | null | null | null | null | null | null | [] | null | null | null | fp8 | null | null | false | false | rohit | false | Abliterated model with MTP (3 speculative tokens), FP8 KV cache. Decode measured with 1000-token output (MTP steady-state). TTFT from warm prefix cache. GPU running at 600W max TDP. 600w |
external-community:cmofqbobm0016l504s1pd2x5c | external-community | cmofqbobm0016l504s1pd2x5c | 2026-08-10T02:44:31Z | 2026-04-26T12:13:13.570Z | 1 | sakamakismile/Huihui-Qwen3.6-27B-abliterated-NVFP4-MTP | main | Huihui-Qwen3.6-27B-abliterated-NVFP4-MTP | huihui-ai/Huihui-Qwen3.6-27B-abliterated | Qwen | 27 | null | false | vllm | v0.19.2rc1.dev134+gfe9c3d6c5 | cuda | NVFP4 | RTX 5090 | DISCRETE_GPU | RTX 5090 | 1 | 32 | null | null | Ryzen 7 3700X | null | Ubuntu 24.04 | 21 | 1,000 | 131,072 | 1 | null | 112 | 94.4 | null | null | null | null | null | null | [] | null | null | null | fp8 | null | true | false | false | rohit | false | Abliterated model with MTP (3 speculative tokens), FP8 KV cache. Decode measured with 1000-token output (MTP steady-state). TTFT from warm prefix cache. GPU running at 400W max TDP. Average of 6 runs. |
external-community:cmofvuf7v0005jv046xvcrah0 | external-community | cmofvuf7v0005jv046xvcrah0 | 2026-08-10T02:44:31Z | 2026-04-26T14:47:46.315Z | 1 | Qwen/Qwen2.5-7B-Instruct | main | Qwen2.5-7B-Instruct | Qwen/Qwen2.5-7B | Qwen | 7 | null | false | llama.cpp | null | cuda | Q4_K_M | RTX 4070 Super | DISCRETE_GPU | RTX 4070 Super | 1 | 12 | null | null | AMD Ryzen (CachyOS) | null | CachyOS Linux | 0 | 0 | 65,536 | 1 | null | 248 | 77.4 | null | 85 | null | null | null | 7 | [] | null | null | 99 | turbo2 | null | true | false | false | leclampin | false | Qwen3.5-9B-Q4_K_M (nouveau modèle non encore sur HF) - Métriques réelles depuis logs - RTX 4070 SUPER 12GB - TurboQuant turbo2 |
external-community:cmofyatrj0002jv04r8064iqg | external-community | cmofyatrj0002jv04r8064iqg | 2026-08-10T02:44:31Z | 2026-04-26T15:56:30.896Z | 1 | Qwen/Qwen3.5-27B | main | Qwen3.5-27B | null | Qwen | 28 | null | false | hipfire | 0.1.8-alpha+3945bb2 | rocm | MQ4 | RX 7900 XTX | DISCRETE_GPU | RX 7900 XTX | 1 | 24 | null | null | null | null | null | 27 | 157 | 4,096 | 1 | null | 154.7 | 250.25 | null | null | null | null | null | 18.289063 | [] | null | null | null | null | null | null | true | false | schuttdev | false | hipfire @ 3945bb2 (master post-PR #51 loop-break + ngram_block series)
prompt: merge_sort thinking-OFF (md5=253c7ac50857fe6d0e10fb0d2c5e35c0)
chatml-wrapped + explicit empty <think></think> for thinking-off
prompt file: benchmarks/prompts/merge_sort_thinking_off.txt
runs: 3 (median reported); range 249.6–251.0
pe... |
external-community:cmofzk74w0002ij042txvvkx5 | external-community | cmofzk74w0002ij042txvvkx5 | 2026-08-10T02:44:31Z | 2026-04-26T16:31:47.744Z | 1 | Qwen/Qwen3.5-9B | main | Qwen3.5-9B | Qwen/Qwen3.5-9B-Base | Qwen | 9 | null | false | hipfire | 0.1.8-alpha+e659452 | rocm | MQ4 | RX 7900 XTX | DISCRETE_GPU | RX 7900 XTX | 1 | 24 | null | null | null | null | null | 27 | 184 | 4,096 | 1 | null | 75.53 | 575.25 | null | null | null | null | null | 8.074219 | [] | null | null | null | null | null | null | true | false | schuttdev | false | hipfire @ e659452 (master post-PR #51 + #52 series)
prompt: merge_sort thinking-OFF (md5=253c7ac50857fe6d0e10fb0d2c5e35c0)
chatml-wrapped + explicit empty <think></think> for thinking-off
prompt file: benchmarks/prompts/merge_sort_thinking_off.txt
runs: 3 (median reported); range 571.2–576.0
per-run tok/s: [571.2... |
external-community:cmogixabp0004kw0453ux7bjd | external-community | cmogixabp0004kw0453ux7bjd | 2026-08-10T02:44:31Z | 2026-04-27T01:33:51.109Z | 1 | Qwen/Qwen3.6-35B-A3B | main | Qwen3.6-35B-A3B | null | Qwen | 36 | 3 | true | vllm | 0.19.2rc1.dev205+g07351e088 | null | INT4 | RTX 3090 | DISCRETE_GPU | RTX 3090 | 2 | 24 | null | null | Ryzen | null | windows | 22 | 1,000 | 32,768 | 1 | null | 122 | 140.06 | null | null | null | null | null | 21.05 | [] | null | 2 | null | fp16 | null | null | true | false | wasifb | false | Quant: Intel/Qwen3.6-35B-A3B-Int4-AutoRound (HF gated, base hfId used). Genesis v7.48 (P67 multi-query Triton kernel + P67b spec-verify routing + P78 capture-guard). DFlash N=4 is the throughput sweet spot of an N-sweep on this stack (N=3 134, N=4 140, N=5 137 TPS). Mean acceptance length 2.44, avg draft accept 36.0%, ... |
external-community:cmogj4tyd0001ic0446kgjl3h | external-community | cmogj4tyd0001ic0446kgjl3h | 2026-08-10T02:44:31Z | 2026-04-27T01:39:43.141Z | 1 | Qwen/Qwen3.6-35B-A3B | main | Qwen3.6-35B-A3B | null | Qwen | 36 | 3 | true | vllm | 0.19.2rc1.dev205+g07351e088 | null | INT4 | RTX 3090 | DISCRETE_GPU | RTX 3090 | 2 | 24 | null | null | Ryzen | null | windows | 22 | 1,000 | 32,768 | 1 | null | 120 | 134.4 | null | null | null | null | null | 21 | [] | null | 2 | null | fp16 | null | null | true | false | wasifb | false | Quant: Intel/Qwen3.6-35B-A3B-Int4-AutoRound (HF gated, base hfId used). Genesis v7.48 (P67/P67b/P78). DFlash N=3 datapoint of an N-sweep (N=3 134, N=4 140, N=5 137 TPS). N=3 has highest avg draft accept rate (43.9%) but lowest AL (2.32) → less throughput than N=4. Per-position accept 0.71 / 0.42 / 0.21. Same hardware/p... |
external-community:cmogjbets0008kw045ktjd4zx | external-community | cmogjbets0008kw045ktjd4zx | 2026-08-10T02:44:31Z | 2026-04-27T01:44:50.128Z | 1 | unsloth/Qwen3.6-35B-A3B-GGUF | main | Qwen3.6-35B-A3B-GGUF | Qwen/Qwen3.6-35B-A3B | Qwen | 35 | null | false | llama.cpp | Tom-TurboQuant-fork-b3401eaf9 | null | Q3_K_XL | RTX 3090 | DISCRETE_GPU | RTX 3090 | 2 | 24 | null | null | Ryzen | null | windows | 22 | 1,000 | 131,072 | 1 | null | null | 125.94 | null | null | null | null | null | 10.1 | [] | null | null | 999 | q8_0/turbo4 (asymmetric) | null | true | false | false | wasifb | false | Tom Turney's TurboQuant+ fork of llama.cpp (commit b3401eaf9). Asymmetric KV: q8_0 keys + turbo4 values per @no_stp_on_snek's twitter post. 2× RTX 3090 layer-split, 50/50 tensor-split. Replication of his 2× GTX 1080 Ti benchmark — Pascal got 36 TPS at 128K, Ampere gets 125.94 TPS = 3.5× faster on same model + KV scheme... |
external-community:cmogm5x010002jv040zuze5n3 | external-community | cmogm5x010002jv040zuze5n3 | 2026-08-10T02:44:31Z | 2026-04-27T03:04:32.594Z | 1 | unsloth/Qwen3.6-27B-GGUF | main | Qwen3.6-27B-GGUF | Qwen/Qwen3.6-27B | Qwen | 27 | null | false | llama.cpp | b8778 | rocm | Q3_K_XL | RX 7900 XTX | DISCRETE_GPU | RX 7900 XTX | 1 | 24 | null | null | null | null | null | 512 | 128 | 640 | 1 | null | null | 33.31 | null | null | null | null | null | 13.47 | [] | null | null | null | q8_0 | null | true | false | false | itgoatee | false | llama-bench on ROCm (hip-radeon); tg128 metric; b=1024/ub=512; q8_0/q8_0 KV; gen perfectly flat tg128-tg2048 (33.31->33.08 t/s, -0.7%); DeltaNet hybrid arch; b=1024/ub=512 beats b=2048/ub=512 on prefill (951 vs 943 pp512) |
external-community:cmogmlw6j0006ky04rs6r76ob | external-community | cmogmlw6j0006ky04rs6r76ob | 2026-08-10T02:44:31Z | 2026-04-27T03:16:58.028Z | 1 | GestaltLabs/Ornstein-Hermes-3.6-27b-MLX-8bit | main | Ornstein-Hermes-3.6-27b-MLX-8bit | GestaltLabs/Ornstein-Hermes-3.6-27b | Qwen | 27 | null | false | mlx | null | null | 8bit | M4 Max | UNIFIED | null | 1 | null | 128 | null | null | null | null | 0 | 16 | 2,048 | 1 | null | null | 18 | null | null | null | null | null | null | [] | null | null | null | null | null | null | false | false | hornsan1 | false | mlx_vlm.server, post-warmup short-burst peak. 256-tok sustained 16.3 tok/s. Qwen3.6 arch, 8-bit affine group_size 64. |
external-community:cmogo4m0y0004l204qlx8n2nt | external-community | cmogo4m0y0004l204qlx8n2nt | 2026-08-10T02:44:31Z | 2026-04-27T03:59:30.947Z | 1 | mlx-community/DeepSeek-V4-Flash-2bit-DQ | main | DeepSeek-V4-Flash-2bit-DQ | null | Deepseek | 284 | null | true | mlx | null | null | 2bit-DQ | M4 Max | UNIFIED | null | 1 | null | 128 | null | null | null | null | 0 | 256 | 2,048 | 1 | null | null | 17 | null | null | null | null | null | null | [] | null | null | null | null | null | null | false | false | hornsan1 | false | mlx-lm direct (sustained, 256-tok runs). 284B/13B-active MoE, mixed-precision DQ (2-bit experts, 4-6-8 bit sensitive layers). Required mlx-lm PR #1201 (akashgoswami/add-deepseek-v4) plus local sanitize patches to map HF checkpoint naming -> PR module names and split wo_a [8*1024,4096] across 8 group Linears, plus quant... |
external-community:cmogpa6m10002k004199sqmb4 | external-community | cmogpa6m10002k004199sqmb4 | 2026-08-10T02:44:31Z | 2026-04-27T04:31:50.521Z | 1 | unsloth/Qwen3.6-27B-GGUF | 82d411acf4a06cfb8d9b073a5211bf410bfc29bf | Qwen3.6-27B-GGUF | Qwen/Qwen3.6-27B | Qwen | 27 | null | false | llama.cpp | b8945-30759dfa0 | cuda | UD-Q4_K_XL | RTX 3090 | DISCRETE_GPU | RTX 3090 | 1 | 24 | null | null | null | null | null | 30 | 500 | 2,048 | 1 | null | null | 150.073311 | null | 154.707 | null | null | null | 19.525 | [] | null | 1 | 99 | null | flash_attn | true | true | false | joaosump | false | DFlash speculative decoding benchmark submitted by Codex after two warmup requests against llama-server /v1/chat/completions. Recorded request used temperature=0, max_tokens=500, prompt: Write a Python binary search tree implementation with insert, search, delete, and traversal methods. DFlash acceptance on recorded ru... |
external-community:cmogqkco20007l704rhj8sq83 | external-community | cmogqkco20007l704rhj8sq83 | 2026-08-10T02:44:31Z | 2026-04-27T05:07:44.546Z | 1 | Qwen/Qwen3.6-27B | main | Qwen3.6-27B | null | Qwen | 28 | null | false | llama.cpp | lucebox-hub b6ffab4a9 | cuda | Q4_K_M | RTX 3090 Ti | DISCRETE_GPU | RTX 3090 Ti | 1 | 24 | null | null | AMD Ryzen 9 9950X | null | Ubuntu 24.04 | 117 | 512 | 1,024 | 1 | null | 710 | 140.28 | null | 146.6 | null | null | null | 22 | [] | null | null | null | f16 | null | true | true | false | Skiipy | false | DFlash speculative decoding via Lucebox-Hub test_dflash binary (custom C++/CUDA on top of Luce-Org/llama.cpp@luce-dflash ggml fork). Block-diffusion draft from z-lab/Qwen3.6-27B-DFlash (BF16 safetensors, 3.46 GB, 2026-04-26 z-lab snapshot). DDTree tree-verify budget=22, chain_seed=1, fa_window=2048, KV f16. Bench: scri... |
external-community:cmogsauq70006k40418jb5vob | external-community | cmogsauq70006k40418jb5vob | 2026-08-10T02:44:31Z | 2026-04-27T05:56:20.624Z | 1 | Qwen/Qwen3.6-35B-A3B | main | Qwen3.6-35B-A3B | null | Qwen | 36 | 3 | true | llama.cpp | b1-6217b49 (CUDA) | cuda | UD-Q6_K_XL | RTX 3090 Ti | DISCRETE_GPU | RTX 3090 Ti | 2 | 24 | null | null | AMD Ryzen 9 9950X | null | Ubuntu 24.04 | 512 | 512 | 65,536 | 1 | null | 212.55 | 131.58 | null | 263.17 | null | null | null | 17 | [] | null | null | 999 | q8_0 | null | true | false | false | Skiipy | false | Qwen 3.6 35B-A3B MoE (3B active) UD-Q6_K_XL via llama-server (Unsloth Dynamic 2.0). Dual RTX 3090 (Ti slot 1 Gen4 x16 + Dell slot 2 Gen4 x4), default layer split. Bench: vllm bench serve openai-chat random 512/512 num-prompts 20 concurrency 1 --ignore-eos. TPOT 7.20 ms — MoE 3B-active sparse activation gives sharply lo... |
external-community:cmogsu1xc000mjv04iforumvd | external-community | cmogsu1xc000mjv04iforumvd | 2026-08-10T02:44:31Z | 2026-04-27T06:11:16.416Z | 1 | Qwen/Qwen3-8B | main | Qwen3-8B | Qwen/Qwen3-8B-Base | Qwen | 8 | null | false | llama.cpp | b1-6217b49 (CUDA) | cuda | Q4_K_M | RTX 3090 Ti | DISCRETE_GPU | RTX 3090 Ti | 1 | 24 | null | null | AMD Ryzen 9 9950X | null | Ubuntu 24.04 | 512 | 512 | 8,192 | 1 | null | 122.57 | 133.01 | null | 266.01 | null | null | null | 7 | [] | null | null | 999 | q8_0 | null | true | false | false | Skiipy | false | Qwen 3-8B Q4_K_M via llama-server on single RTX 3090 Ti (slot 1 Gen4 x16). Bench: vllm bench serve openai-chat random 512/512 num-prompts 20 concurrency 1 --ignore-eos. TPOT 7.29 ms, ITL 7.34 ms — 8B Q4 is HBM-bandwidth-saturated on 3090 Ti at ~135 t/s decode. KV q8_0, ngl 999, fa on, no spec decode. ~5 GB on disk, ~7 ... |
external-community:cmogt0ksa000wjv04oqlbcm4w | external-community | cmogt0ksa000wjv04oqlbcm4w | 2026-08-10T02:44:31Z | 2026-04-27T06:16:20.795Z | 1 | Qwen/Qwen3.5-9B | main | Qwen3.5-9B | Qwen/Qwen3.5-9B-Base | Qwen | 9 | null | false | llama.cpp | b1-6217b49 (CUDA) | cuda | UD-Q4_K_XL | RTX 3090 Ti | DISCRETE_GPU | RTX 3090 Ti | 1 | 24 | null | null | AMD Ryzen 9 9950X | null | Ubuntu 24.04 | 512 | 512 | 8,192 | 1 | null | 155.81 | 119.05 | null | 238.09 | null | null | null | 8 | [] | null | null | 999 | q8_0 | null | true | false | false | Skiipy | false | Qwen 3.5-9B unsloth Dynamic 2.0 UD-Q4_K_XL via llama-server on single RTX 3090 Ti (slot 1 Gen4 x16). Bench: vllm bench serve openai-chat random 512/512 num-prompts 20 concurrency 1 --ignore-eos. TPOT 8.11 ms. KV q8_0, ngl 999, fa on, no spec decode. ~6 GB on disk (5.96 GB), ~8 GB peak VRAM. Model is from Qwen 3.5 Small... |
external-community:cmoh6rfig000mju04hf34f6re | external-community | cmoh6rfig000mju04hf34f6re | 2026-08-10T02:44:31Z | 2026-04-27T12:41:08.681Z | 1 | Brooooooklyn/Qwen3.6-35B-A3B-UD-Q8_K_XL-mlx | 3f8ff65551b3bfda2d8ccfadc091bab9ce066905 | Qwen3.6-35B-A3B-UD-Q8_K_XL-mlx | Qwen/Qwen3.6-35B-A3B | Qwen | 35 | 3 | true | mlx | @mlx-node/lm 0.0.7 | metal | Q8_K_XL | M4 Max | UNIFIED | null | 1 | null | 64 | null | Apple M4 Max 16-core CPU / 40-core GPU | null | macOS 26.4 build 25E246 | 0 | 512 | 262,144 | 1 | null | 105.75 | 83.353 | null | null | null | null | null | 22 | [] | null | null | null | null | null | null | false | false | shikhar | false | Automated local benchmark on Mac Studio. Runner: Node v25.9.0 + @mlx-node/lm 0.0.7. Command was run from ~/mlx-node-qwen36-test. Submitted tokSOut/ttftMs use warm runs 2-3 average; run 1 is included below as post-load warmup. Raw output: run=1 ttft_ms=1572.6 elapsed_s=7.419 tokens=509 tok_s=68.608; run=2 ttft_ms=118.8 ... |
external-community:cmoh721ve000kla04zc0ki9gs | external-community | cmoh721ve000kla04zc0ki9gs | 2026-08-10T02:44:31Z | 2026-04-27T12:49:24.218Z | 1 | unsloth/Qwen3.6-35B-A3B | main | Qwen3.6-35B-A3B | Qwen/Qwen3.6-35B-A3B | Qwen | 36 | null | false | llama.cpp | b8778 | rocm | Q3_K_XL | RX 7900 XTX | DISCRETE_GPU | RX 7900 XTX | 1 | 24 | null | null | null | null | null | 512 | 128 | 640 | 1 | null | null | 104.53 | null | null | null | null | null | 15.68 | [] | null | null | null | q8_0 | null | true | false | false | itgoatee | false | llama-bench on ROCm (hip-radeon); tg128 metric; b/ub 1024; q8_0/q8_0 KV; all-GPU no -ot needed (15.68 GiB fits 24 GB VRAM); gen flat 104-106 t/s tg128-tg2048; b/ub 1024 wins prefill over b/ub 512 (pp4096: 3035 vs 2353); b/ub 2048 catastrophic cliff (pp4096=653, -78%) |
external-community:cmohdd82h000okw04we7c4ggp | external-community | cmohdd82h000okw04we7c4ggp | 2026-08-10T02:44:31Z | 2026-04-27T15:46:03.161Z | 1 | llmfan46/Qwen3.6-35B-A3B-uncensored-heretic | main | Qwen3.6-35B-A3B-uncensored-heretic | Qwen/Qwen3.6-35B-A3B | Qwen | 36 | 3 | true | llama.cpp | b8931 | cuda | Q3_K_M | RTX 4090 | DISCRETE_GPU | RTX 4090 | 1 | 24 | null | null | Intel Core Ultra 9 285K | null | Fedora 43 | 0 | 512 | 8,192 | 1 | null | 183.04 | 164.09 | null | null | null | null | null | null | [] | null | null | 999 | q8_0 | null | true | false | false | Bradarr | false | Median of 3 runs (164.09, 166.02, 163.92 tok/s). Thinking disabled for pure decode benchmark. Router-based multi-model setup via llama-b8931 models.ini. Heretic v1.2.0 uncensored finetune. |
external-community:cmohdl6vt000wkw04nd6tci35 | external-community | cmohdl6vt000wkw04nd6tci35 | 2026-08-10T02:44:31Z | 2026-04-27T15:52:14.873Z | 1 | Qwen/Qwen3.6-35B-A3B | main | Qwen3.6-35B-A3B | null | Qwen | 36 | 3 | true | llama.cpp | b8931 | cuda | Q3_K_P | RTX 4090 | DISCRETE_GPU | RTX 4090 | 1 | 24 | null | null | Intel Core Ultra 9 285K | null | Fedora 43 | 0 | 512 | 8,192 | 1 | null | 174.91 | 157.72 | null | null | null | null | null | null | [] | null | null | 999 | q8_0 | null | true | false | false | Bradarr | false | HauhauCS Aggressive uncensored finetune (GGUF-only repo, submitted under base model). Median of 3 runs (165.26, 157.72, 157.66). Thinking disabled. Router-based llama-b8931. |
external-community:cmohdrtfo000hkw04cjyedef1 | external-community | cmohdrtfo000hkw04cjyedef1 | 2026-08-10T02:44:31Z | 2026-04-27T15:57:24.037Z | 1 | Qwen/Qwen3.6-35B-A3B | main | Qwen3.6-35B-A3B | null | Qwen | 36 | 3 | true | llama.cpp | b8931 | cuda | Q3_K_P | RTX 4090 | DISCRETE_GPU | RTX 4090 | 1 | 24 | null | null | Intel Core Ultra 9 285K | null | Fedora 43 | 0 | 512 | 8,192 | 1 | null | 174.91 | 157.72 | null | null | null | null | null | null | [] | null | null | 999 | q8_0 | null | true | false | false | Bradarr | false | HauhauCS Aggressive uncensored finetune (GGUF-only repo, submitted under base model). Median of 3 runs (165.26, 157.72, 157.66). Thinking disabled. Router-based llama-b8931. |
external-community:cmohe2eo50019kw04s0yxiow1 | external-community | cmohe2eo50019kw04s0yxiow1 | 2026-08-10T02:44:31Z | 2026-04-27T16:05:38.118Z | 1 | Qwen/Qwen3.6-27B | main | Qwen3.6-27B | null | Qwen | 28 | null | false | llama.cpp | b8931 | cuda | Q4_K_P | RTX 4090 | DISCRETE_GPU | RTX 4090 | 1 | 24 | null | null | Intel Core Ultra 9 285K | null | Fedora 43 | 0 | 512 | 8,192 | 1 | null | 151.22 | 42.7 | null | null | null | null | null | null | [] | null | null | 999 | q8_0 | null | true | false | false | Bradarr | false | HauhauCS Aggressive uncensored finetune (GGUF-only repo, submitted under base model). Median of 3 runs (42.84, 42.42, 42.70). Thinking disabled. Router-based llama-b8931. |
external-community:cmohe99tn0015js04kxz8g21x | external-community | cmohe99tn0015js04kxz8g21x | 2026-08-10T02:44:31Z | 2026-04-27T16:10:58.427Z | 1 | unsloth/Qwen3.6-27B | main | Qwen3.6-27B | Qwen/Qwen3.6-27B | Qwen | 28 | null | false | llama.cpp | b8931 | cuda | Q3_K_XL | RTX 4090 | DISCRETE_GPU | RTX 4090 | 1 | 24 | null | null | Intel Core Ultra 9 285K | null | Fedora 43 | 0 | 512 | 8,192 | 1 | null | 162.51 | 49.37 | null | null | null | null | null | null | [] | null | null | 999 | q8_0 | null | true | false | false | Bradarr | false | unsloth UD dynamic quant. Median of 3 runs (49.37, 49.00, 49.40). Thinking disabled. Router-based llama-b8931. |
external-community:cmohg8foe001hkw048rd58ig8 | external-community | cmohg8foe001hkw048rd58ig8 | 2026-08-10T02:44:31Z | 2026-04-27T17:06:18.590Z | 1 | TurbulenceDeterministe/Carnice-9b-W8A16-AWQ | main | Carnice-9b-W8A16-AWQ | kai-os/Carnice-9b | Qwen | 9 | null | false | vllm | null | null | AWQ | RTX 3090 | DISCRETE_GPU | RTX 3090 | 2 | 24 | null | null | AMD EPYC 7543 | null | Ubuntu 24.04 | 1,024 | 256 | 2,048 | 1 | null | 216 | 114.26 | null | 571.31 | null | null | null | 22.7 | [] | null | null | null | null | null | null | false | false | TurbulenceDeterministe | false | Single-user throughput. concurrency=1, 50 prompts, 1024 in/256 out tokens. Mean TPOT 7.88ms, median TTFT 219ms, P99 TTFT 222ms. Served via llama-swap proxy. Peak output tok/s: 128. |
external-community:cmohkdafg000nl704l43fxf7j | external-community | cmohkdafg000nl704l43fxf7j | 2026-08-10T02:44:31Z | 2026-04-27T19:02:03.532Z | 1 | unsloth/Qwen3.6-27B | main | Qwen3.6-27B | Qwen/Qwen3.6-27B | Qwen | 28 | null | false | llama.cpp | buun-llama-cpp b8959-7343d30c6 | cuda | UD-Q4_K_XL | RTX 3090 Ti | DISCRETE_GPU | RTX 3090 Ti | 1 | 24 | null | null | null | null | null | 30 | 499 | 4,096 | 1 | null | 67.56 | 155.9375 | null | 165.3125 | null | null | null | 18.4 | [] | null | 1 | 99 | null | flash_attn | true | true | false | Skiipy | false | DFlash speculative decoding bench reproducing joaosump's methodology: 2 warmup requests + recorded request against llama-server /v1/chat/completions. Recorded request used temperature=0, max_tokens=500, prompt: Write a Python binary search tree implementation with insert, search, delete, and traversal methods. DFlash a... |
external-community:cmohodksg000ejp04g9tofth9 | external-community | cmohodksg000ejp04g9tofth9 | 2026-08-10T02:44:31Z | 2026-04-27T20:54:15.425Z | 1 | Qwen/Qwen3.6-35B-A3B | main | Qwen3.6-35B-A3B | null | Qwen | 36 | 3 | true | ollama | null | null | Q4_K_M | RTX 5090 | DISCRETE_GPU | RTX 5090 | 1 | 32 | null | null | Ryzen | null | null | 535 | 402 | 32,768 | 1 | null | 0.5 | 175.8 | null | null | null | null | null | 26.1 | [] | null | null | 99 | null | null | true | false | false | rubysuelocal-rgb | false | null |
external-community:cmoiu2lfb0012l804luvjy1m2 | external-community | cmoiu2lfb0012l804luvjy1m2 | 2026-08-10T02:44:31Z | 2026-04-28T16:21:26.904Z | 1 | sakamakismile/Huihui-Qwen3.6-27B-abliterated-NVFP4-MTP | main | Huihui-Qwen3.6-27B-abliterated-NVFP4-MTP | huihui-ai/Huihui-Qwen3.6-27B-abliterated | Qwen | 27 | null | false | vllm | null | null | NVFP4 | RTX 5090 | DISCRETE_GPU | RTX 5090 | 1 | 32 | null | null | Ryzen 7 3700X | null | Ubuntu 24.04 | 97,100 | 1,000 | 131,072 | 1 | null | 115 | 97.5 | null | null | null | null | null | 32 | [] | null | null | null | null | null | null | false | false | rohit | false | MTP active (num_speculative_tokens=3). Abliterated model. KV cache dtype: fp8. --enable-chunked-prefill enabled. GPU power capped at 450W. Max model length: 131072. |
external-community:cmoj30q1i000al704r6kdoz7d | external-community | cmoj30q1i000al704r6kdoz7d | 2026-08-10T02:44:31Z | 2026-04-28T20:31:56.118Z | 1 | unsloth/Qwen3.6-27B | main | Qwen3.6-27B | Qwen/Qwen3.6-27B | Qwen | 28 | null | false | lmstudio | null | metal | Q4_K_M | M4 Max | UNIFIED | null | 1 | null | 48 | null | null | null | 26.1 | 43 | 700 | 4,096 | 1 | null | 0.83 | 21.84 | null | null | null | null | null | null | [] | null | null | 64 | null | null | true | false | false | KovD3v | false | null |
external-community:cmok73fys000dl1040u12pav9 | external-community | cmok73fys000dl1040u12pav9 | 2026-08-10T02:44:31Z | 2026-04-29T15:13:47.668Z | 1 | Qwen/Qwen3.6-27B | main | Qwen3.6-27B | null | Qwen | 28 | null | false | llama.cpp | b8963 | cuda | IQ4_NL | RTX 4090 | DISCRETE_GPU | RTX 4090 | 1 | 24 | null | null | Intel Core Ultra 9 285K | null | Fedora 43 | 252 | 1,024 | 98,304 | 1 | null | 850 | 44.2 | null | 55.1 | null | null | null | null | [] | null | null | 99 | q8_0 | null | true | false | false | YourHighnessLA | false | Automated benchmark. 252-token prompt + 1024 output tokens. RTX 4090 (24GB), Ryzen 7 7800X3D, WSL2. |
external-community:cmokpfzeh0007l104s2fb4qm9 | external-community | cmokpfzeh0007l104s2fb4qm9 | 2026-08-10T02:44:31Z | 2026-04-29T23:47:25.817Z | 1 | sakamakismile/Qwen3.6-27B-Text-NVFP4-MTP | main | Qwen3.6-27B-Text-NVFP4-MTP | Qwen/Qwen3.6-27B | Qwen | 17 | null | false | vllm | 0.20.0 | cuda | NVFP4 | RTX 5090 | DISCRETE_GPU | RTX 5090 | 1 | 32 | null | null | Ryzen 7 3700X | null | Ubuntu 24.04 | 512 | 512 | 114,688 | 1 | null | 82.99 | 101.87 | null | 203.74 | null | null | null | 30 | [] | null | null | null | bf16 | null | true | true | false | Skiipy | false | Single RTX 5090 ROG Astral OC (Blackwell GB202, sm_120a), NVFP4 modelopt 0.43.0 checkpoint with MTP head restored in bf16. vLLM 0.20.0 + flashinfer-python 0.6.8.post1 + flashinfer-cubin 0.6.8.post1 + torch 2.11.0 + CUDA toolkit 12.9. MTP n=3 acceptance 60.20% (per-position 74.27/59.13/47.19), accept length 2.81. bf16 K... |
external-community:cmokq33sw0004ks0413sul32a | external-community | cmokq33sw0004ks0413sul32a | 2026-08-10T02:44:31Z | 2026-04-30T00:05:24.608Z | 1 | z-lab/Qwen3.6-27B-DFlash | main | Qwen3.6-27B-DFlash | null | Qwen | 2 | null | false | llama.cpp | lucebox-hub b6ffab4a9 (sm_120 rebuild) | cuda | Q4_K_M | RTX 5090 | DISCRETE_GPU | RTX 5090 | 1 | 32 | null | null | Ryzen 7 3700X | null | Ubuntu 24.04 | 117 | 512 | 1,024 | 1 | null | 470 | 215.43 | null | 225.13 | null | null | null | 23 | [] | null | null | null | f16 | null | true | true | false | Skiipy | false | DFlash speculative decoding via Lucebox-Hub test_dflash binary on RTX 5090 ROG Astral OC (Blackwell GB202). Same bench methodology as prior #19 submission (3090 Ti, 140.28 t/s mean) — scripts/bench_he.py --n-gen 512 --ddtree-budget 22 over 10 HumanEval prompts, single-stream batch=1. 5090 result: mean 215.43 t/s (range... |
external-community:cmolpeunj000el2045o540qlc | external-community | cmolpeunj000el2045o540qlc | 2026-08-10T02:44:31Z | 2026-04-30T16:34:19.184Z | 1 | Qwen/Qwen3.6-35B-A3B | main | Qwen3.6-35B-A3B | null | Qwen | 36 | 3 | true | llama.cpp | b8986 | cuda | IQ3_XXS | RTX 5080 | DISCRETE_GPU | RTX 5080 | 1 | 16 | null | null | null | null | Debian 13.4 (WSL2) | 512 | 512 | 131,072 | 1 | null | null | 150.6 | null | 1,026.4 | null | null | null | 14.2 | [] | null | null | 99 | null | null | true | false | false | leftcurvedev | false | OS/Screen already take 0.8-1.2GB VRAM
Total usage with everything (model + OS): ~15.2GB
|
external-community:cmolvxd82000alb04jarmcutc | external-community | cmolvxd82000alb04jarmcutc | 2026-08-10T02:44:31Z | 2026-04-30T19:36:40.754Z | 1 | Qwen/Qwen3.6-27B | main | Qwen3.6-27B | null | Qwen | 28 | null | false | llama.cpp | b8818-b572d1ecd | cuda | Q4_K_M | RTX 4090 | DISCRETE_GPU | RTX 4090 | 1 | 24 | null | null | Intel Core Ultra 9 285K | null | Fedora 43 | 512 | 512 | 90,200 | 2,048 | null | 409 | 46.64 | null | 45.31 | null | null | null | 18.3 | [] | null | null | 99 | q8_0 | null | true | false | false | adybag14-cyber | false | RTX 4090 24GB, WSL2, CUDA toolkit 12.8, Windows NVIDIA driver 591.59. llama.cpp build b8818-b572d1ecd, CUDA arch 89. Command used -ngl 99 -c 90200 -b 2048 -ub 512 --flash-attn on --cache-type-k q8_0 --cache-type-v q8_0 -np 1 --jinja --no-mmap. Model fully offloaded: 65/65 layers. |
external-community:cmolw295o000tju04uiwr820b | external-community | cmolw295o000tju04uiwr820b | 2026-08-10T02:44:31Z | 2026-04-30T19:40:28.764Z | 1 | Qwen/Qwen3.6-35B-A3B | main | Qwen3.6-35B-A3B | null | Qwen | 36 | 3 | true | llama.cpp | null | rocm | IQ4_XS | RX 7900 XTX | DISCRETE_GPU | RX 7900 XTX | 1 | 24 | null | null | null | null | null | 512 | 512 | 2,048 | 4,096 | null | 59 | 85.6 | null | null | null | null | null | null | [] | null | null | 99 | null | null | true | false | false | veladan | false | null |
external-community:cmonueb7t0002kv04rhpfu8jb | external-community | cmonueb7t0002kv04rhpfu8jb | 2026-08-10T02:44:31Z | 2026-05-02T04:29:24.426Z | 1 | Qwen/Qwen3.5-9B | main | Qwen3.5-9B | Qwen/Qwen3.5-9B-Base | Qwen | 9 | null | false | llama.cpp | null | cuda | Q8_0 | RTX 3080 Ti | DISCRETE_GPU | RTX 3080 Ti | 1 | 12 | null | null | N/A | null | Linux | 0 | 0 | 250,000 | 1 | null | 164,149.51 | 15.36 | null | null | null | null | null | null | [] | null | null | null | null | null | null | false | false | aam2bench | false | Benchmarked Unsloth/Qwen3.5-9B-Q8_0.gguf with tg32 test at d250000 context on RTX 3080 Ti. Achieved 15.36 t/s with 16.00 peak t/s. |
external-community:cmonuslgg0008kv046wdb0suf | external-community | cmonuslgg0008kv046wdb0suf | 2026-08-10T02:44:31Z | 2026-05-02T04:40:30.880Z | 1 | unsloth/Qwen3.6-35B-A3B | main | Qwen3.6-35B-A3B | Qwen/Qwen3.6-35B-A3B | Qwen | 36 | null | false | llama.cpp | 27aef3dd9 | cuda | Q4_K_M | RTX 4070 | DISCRETE_GPU | RTX 4070 | 1 | 12 | null | null | null | null | Ubuntu 22.04 | 216 | 512 | 2,048 | 1 | null | 68 | 54.5 | null | 53 | null | null | null | null | [] | null | null | null | null | null | null | false | false | iam-shanmukha | false | null |
external-community:cmonv2bj40001ju04qxmz2wg4 | external-community | cmonv2bj40001ju04qxmz2wg4 | 2026-08-10T02:44:31Z | 2026-05-02T04:48:04.576Z | 1 | unsloth/Qwen3.6-35B-A3B | main | Qwen3.6-35B-A3B | Qwen/Qwen3.6-35B-A3B | Qwen | 36 | null | false | llama.cpp | 27aef3dd9 | cuda | Q4_K_M | RTX 4070 | DISCRETE_GPU | RTX 4070 | 1 | 12 | null | null | null | null | Ubuntu 22.04 | 216 | 512 | 131,072 | 1 | null | 68 | 54.5 | null | 53 | null | null | null | null | [] | null | null | null | null | null | null | false | false | iam-shanmukha | false | null |
external-community:cmonvhnkb0006js046k76k13h | external-community | cmonvhnkb0006js046k76k13h | 2026-08-10T02:44:31Z | 2026-05-02T05:00:00.012Z | 1 | unsloth/Qwen3.6-35B-A3B | main | Qwen3.6-35B-A3B | Qwen/Qwen3.6-35B-A3B | Qwen | 36 | null | false | llama.cpp | 27aef3dd9 | cuda | Q4_K_M | RTX 4070 | DISCRETE_GPU | RTX 4070 | 1 | 12 | null | null | null | null | Ubuntu 22.04 | 1,821 | 465 | 131,072 | 1 | null | 50 | 60.2 | null | 284.2 | null | null | null | 11.6 | [] | null | null | null | null | null | null | false | false | iam-shanmukha | false | null |
external-community:cmonvuxv80009jo04pp8xddmv | external-community | cmonvuxv80009jo04pp8xddmv | 2026-08-10T02:44:31Z | 2026-05-02T05:10:19.893Z | 1 | unsloth/Qwen3.6-35B-A3B | main | Qwen3.6-35B-A3B | Qwen/Qwen3.6-35B-A3B | Qwen | 36 | null | false | llama.cpp | 27aef3dd9 | cuda | Q4_K_M | RTX 4070 | DISCRETE_GPU | RTX 4070 | 1 | 12 | null | null | null | null | Ubuntu 22.04 | 140 | 4,083 | 131,072 | 1 | null | 11,784 | 62.1 | null | 56.1 | null | null | null | null | [] | null | null | null | null | null | null | false | false | iam-shanmukha | false | null |
external-community:cmoonkblc002jl104uwnna2b5 | external-community | cmoonkblc002jl104uwnna2b5 | 2026-08-10T02:44:31Z | 2026-05-02T18:05:53.712Z | 1 | google/gemma-4-E2B-it | main | gemma-4-E2B-it | google/gemma-4-E2B | Gemma | 5 | null | false | llama.cpp | null | cpu | Q4_K_M | Qualcomm Snapdragon 888 ARM64 | CPU_ONLY | null | 1 | null | null | null | Qualcomm Snapdragon 888 ARM64 | null | Android | 30 | 1,306 | 1,306 | 1 | null | 10,000 | 6.2 | null | 6.2 | null | null | null | 6 | [] | null | null | null | null | null | null | false | false | Kylej572 | false | null |
external-community:cmopjebwz0004ky04xsxdihvn | external-community | cmopjebwz0004ky04xsxdihvn | 2026-08-10T02:44:31Z | 2026-05-03T08:57:01.907Z | 1 | unsloth/Qwen3.6-27B | main | Qwen3.6-27B | Qwen/Qwen3.6-27B | Qwen | 28 | null | false | llama.cpp | b8931 | cuda | Q4_K_XL | RTX 5090 | DISCRETE_GPU | RTX 5090 | 1 | 32 | null | null | Ryzen | null | null | 1,670 | 514 | 200,192 | 2,048 | null | null | 57.45 | null | null | null | null | null | 26.358 | [] | null | null | 999 | q8_0 | null | true | false | false | sugattor | false | null |
external-community:cmoqakz2k0004kw04w7793b7v | external-community | cmoqakz2k0004kw04w7793b7v | 2026-08-10T02:44:31Z | 2026-05-03T21:38:01.484Z | 1 | nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 | main | NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 | null | null | 32 | 3 | true | ollama | null | cuda | Q4_K_M | RTX 5090 | DISCRETE_GPU | RTX 5090 | 1 | 32 | null | null | Ryzen | null | null | 512 | 512 | 800 | 1 | null | 200 | 285.6 | null | null | null | null | null | null | [] | null | null | null | null | null | null | false | false | Kylej572 | false | null |
external-community:cmoqt7a830006k104or79bpw7 | external-community | cmoqt7a830006k104or79bpw7 | 2026-08-10T02:44:31Z | 2026-05-04T06:19:15.459Z | 1 | dealignai/MiniMax-M2.7-JANGTQ-CRACK | main | MiniMax-M2.7-JANGTQ-CRACK | MiniMaxAI/MiniMax-M2.7 | Minimax | 15 | null | true | mlx | null | null | JANGTQ | M4 Max | UNIFIED | null | 1 | null | 128 | null | null | null | null | 0 | 0 | 2,048 | 1 | null | null | 31 | null | 77 | null | null | null | null | [] | null | null | null | null | null | null | false | false | hornsan1 | false | Median across 168-task vera-eval suite. Generation-only (tokSOut) median 30.80, peak 41.3. End-to-end (tokSTotal) median 76.92. JANGTQ mixed: routed experts 2-bit, attn/shared/embed/lm_head 8-bit. 230B/10B-active MoE. Custom single-worker mlx_lm server (port 8092), temp>=0.7 required. |
external-community:cmoqtkf5e0006ky04t9q5a2fs | external-community | cmoqtkf5e0006ky04t9q5a2fs | 2026-08-10T02:44:31Z | 2026-05-04T06:29:28.371Z | 1 | unsloth/Qwen3.6-27B-UD-MLX-6bit | main | Qwen3.6-27B-UD-MLX-6bit | Qwen/Qwen3.6-27B | Qwen | 11 | null | false | mlx | null | null | 6bit | M4 Max | UNIFIED | null | 1 | null | 128 | null | null | null | null | 0 | 0 | 2,048 | 1 | null | null | 13 | null | 13 | null | null | null | 60.27 | [] | null | null | null | null | null | null | false | false | hornsan1 | false | Median across 169-task vera-eval suite (A grade, 1359.5/1580 — highest-quality dense 27B run). Peak 19.2 tok/s. Unsloth UD (Ultra Dense) 6-bit MLX quant via mlx_vlm.server. |
external-community:cmoqtxiun000gky047ofznawj | external-community | cmoqtxiun000gky047ofznawj | 2026-08-10T02:44:31Z | 2026-05-04T06:39:39.695Z | 1 | prism-ml/Ternary-Bonsai-8B-mlx-2bit | main | Ternary-Bonsai-8B-mlx-2bit | prism-ml/Ternary-Bonsai-8B-unpacked | Qwen | 1 | null | false | mlx | null | null | 2bit-ternary | M4 Max | UNIFIED | null | 1 | null | 128 | null | null | null | null | 0 | 0 | 2,048 | 1 | null | 250 | 135 | 1,100 | null | null | null | null | null | [] | null | null | null | null | null | null | false | false | hornsan1 | false | Manual decode bench: 125-135 tok/s steady (peak 135), sustained 4k generation at 124 tok/s with no degradation. Prefill 1.1k tok/s for long contexts. TTFT 150-360 ms. Long-context 16k: ~13.5s prefill then 125 tok/s decode. Ternary 2-bit MLX quant via mlx_vlm.server. vera-eval: B grade 1244.5/1580. |
external-community:cmotxm0840048qy01oxhgwwt2 | external-community | cmotxm0840048qy01oxhgwwt2 | 2026-08-10T02:44:31Z | 2026-05-06T10:45:59.332Z | 1 | sakamakismile/Huihui-Qwen3.6-27B-abliterated-NVFP4-MTP | main | Huihui-Qwen3.6-27B-abliterated-NVFP4-MTP | huihui-ai/Huihui-Qwen3.6-27B-abliterated | Qwen | 27 | null | false | vllm | 0.19.2rc1.dev134+gfe9c3d6c5 | cuda | NVFP4 | RTX 5090 | DISCRETE_GPU | RTX 5090 | 1 | 32 | null | null | Ryzen 7 3700X | null | Ubuntu 24.04 | 8,379 | 200 | 8,379 | 1 | null | 766 | 105.3 | 10,940 | null | null | null | null | 32 | [] | null | null | null | fp8 | null | null | true | true | rohit | false | Cold benchmark (May 6, 2026). 8K context (8379 prompt tokens → 200 output). Pure gen speed: 105.3 tok/s (temp=0, greedy decode). Prefill throughput: 10940 tok/s. MTP speculative decoding active (num_speculative_tokens=3). Abliterated model. FP8 KV cache, --enable-chunked-prefill, --enable-prefix-caching. GPU power capp... |
external-community:cmoty3p3z004fqy01duetlmcs | external-community | cmoty3p3z004fqy01duetlmcs | 2026-08-10T02:44:31Z | 2026-05-06T10:59:44.736Z | 1 | sakamakismile/Huihui-Qwen3.6-27B-abliterated-NVFP4-MTP | main | Huihui-Qwen3.6-27B-abliterated-NVFP4-MTP | huihui-ai/Huihui-Qwen3.6-27B-abliterated | Qwen | 27 | null | false | vllm | 0.19.2rc1.dev134+gfe9c3d6c5 | cuda | NVFP4 | RTX 5090 | DISCRETE_GPU | RTX 5090 | 1 | 32 | null | null | Ryzen 7 3700X | null | Ubuntu 24.04 | 68,360 | 200 | 68,360 | 1 | null | 12,162 | 100.1 | 5,621 | null | null | null | null | 32 | [] | null | null | null | fp8_e4m3 | null | null | true | true | rohit | false | Cold benchmark (May 6, 2026). ~32K context (68360 prompt tokens → 200 output). Pure gen speed: 100.1 tok/s (temp=0, greedy decode). Prefill throughput: 5621 tok/s. TTFT: 12.16s. MTP speculative decoding active (num_speculative_tokens=3). Abliterated model. FP8 KV cache (fp8_e4m3). --enable-chunked-prefill, --enable-pre... |
external-community:cmovnbp8k006mp101kw5l5jir | external-community | cmovnbp8k006mp101kw5l5jir | 2026-08-10T02:44:31Z | 2026-05-07T15:33:34.724Z | 1 | Qwen/Qwen3.6-35B-A3B | main | Qwen3.6-35B-A3B | null | Qwen | 36 | 3 | true | vllm | 0.19.2rc1.dev220+g7b1bc0a3e | cuda | INT4 | RTX 3090 | DISCRETE_GPU | RTX 3090 | 2 | 24 | null | null | AMD EPYC 7543 | null | Ubuntu 24.04 | 13 | 981 | 32,768 | 1 | null | 136 | 155.57 | null | null | null | null | null | 21.4 | [] | null | 2 | null | fp16 | flash_attn | null | false | false | Mattsav | false | Qwen3.6 35B-A3B INT4 AutoRound using the current available Intel/Qwen3.6-35B-A3B-int4-mixed-AutoRound artifact cached locally. Docker vLLM nightly was faster than the local vLLM 0.19.0 env. Bench: 800-word essay prompt, streaming chat completions, max_tokens=1000, temperature=0.6, top_p=0.95, enable_thinking=false, one... |
external-community:cmovrgyqg009yp101i5t9x0fe | external-community | cmovrgyqg009yp101i5t9x0fe | 2026-08-10T02:44:31Z | 2026-05-07T17:29:38.777Z | 1 | Qwen/Qwen3.6-35B-A3B | main | Qwen3.6-35B-A3B | null | Qwen | 36 | 3 | true | vllm | 0.19.2rc1.dev220+g7b1bc0a3e | cuda | INT4 | RTX 3090 | DISCRETE_GPU | RTX 3090 | 2 | 24 | null | null | AMD EPYC 7543 | null | Ubuntu 24.04 | 13 | 993 | 32,768 | 1 | null | 120 | 167.75 | null | null | null | null | null | 21.3 | [] | null | 2 | null | fp16 | flash_attn | null | false | false | Mattsav | false | Qwen3.6 35B-A3B INT4 AutoRound using public non-mixed mirror abhinand/Qwen3.6-35B-A3B-int4-AutoRound after Intel/Qwen3.6-35B-A3B-int4-AutoRound was unavailable/gated. Docker vLLM nightly TP=2 on GPUs 4-5. Bench: 800-word essay prompt, streaming chat completions, max_tokens=1000, temperature=0.6, top_p=0.95, enable_thin... |
external-community:cmovrue8s00ajp101j30t2qku | external-community | cmovrue8s00ajp101j30t2qku | 2026-08-10T02:44:31Z | 2026-05-07T17:40:05.405Z | 1 | google/gemma-4-E2B-it | main | gemma-4-E2B-it | google/gemma-4-E2B | Gemma | 5 | null | false | llama.cpp | b9033 | cpu | Q4_K_M | Intel R Core TM Ultra 7 155H | CPU_ONLY | null | 1 | null | null | null | Intel(R) Core(TM) Ultra 7 155H | null | Windows 11 Home (build 26200) | 256 | 128 | 2,048 | 1 | null | null | 27.61 | 171.81 | null | null | null | null | null | [] | null | null | null | null | null | null | false | false | dpenev98 | false | Sequential single-benchmark run on Lenovo ThinkPad E16, model 21MA003QBM with Intel(R) Core(TM) Ultra 7 155H, 32 GB RAM, and Windows 11 Home (build 26200). Fixed profile: p=256, n=128, r=3, t=16, b=256, ub=256, ctk=?, ctv=?, mmp=?. Final baseline runs were taken in Windows max-performance mode. CPU baseline for gemma-4... |
external-community:cmovs0uhg00aup1014kmit9lr | external-community | cmovs0uhg00aup1014kmit9lr | 2026-08-10T02:44:31Z | 2026-05-07T17:45:06.389Z | 1 | Qwen/Qwen3.5-4B | main | Qwen3.5-4B | Qwen/Qwen3.5-4B-Base | Qwen | 4 | null | false | llama.cpp | b9033 | cpu | Q4_K_M | Intel R Core TM Ultra 7 155H | CPU_ONLY | null | 1 | null | null | null | Intel(R) Core(TM) Ultra 7 155H | null | Windows 11 Home (build 26200) | 256 | 128 | 2,048 | 1 | null | null | 14.55 | 55.49 | null | null | null | null | null | [] | null | null | null | null | null | null | false | false | dpenev98 | false | Sequential single-benchmark run on a Lenovo ThinkPad E16 (21MA003QBM) with Intel Core Ultra 7 155H, 32 GB RAM, and Windows 11 Home build 26200. Fixed profile: p=256, n=128, r=3, t=16, b=256, ub=256, ctk=q8_0, ctv=f16, mmp=0. Final baseline runs were taken in Windows max-performance mode. CPU result for the winning Qwen... |
external-community:cmovs7b6m00bbp101pr3fb0nn | external-community | cmovs7b6m00bbp101pr3fb0nn | 2026-08-10T02:44:31Z | 2026-05-07T17:50:07.966Z | 1 | Qwen/Qwen3.5-9B | main | Qwen3.5-9B | Qwen/Qwen3.5-9B-Base | Qwen | 9 | null | false | llama.cpp | b9033 | cpu | Q4_K_M | Intel R Core TM Ultra 7 155H | CPU_ONLY | null | 1 | null | null | null | Intel(R) Core(TM) Ultra 7 155H | null | Windows 11 Home (build 26200) | 256 | 128 | 2,048 | 1 | null | null | 8.45 | 29.15 | null | null | null | null | null | [] | null | null | null | null | null | null | false | false | dpenev98 | false | Sequential single-benchmark run on a Lenovo ThinkPad E16 (21MA003QBM) with Intel Core Ultra 7 155H, 32 GB RAM, and Windows 11 Home build 26200. Fixed profile: p=256, n=128, r=3, t=16, b=256, ub=256, ctk=q8_0, ctv=f16, mmp=0. Final baseline runs were taken in Windows max-performance mode. CPU result for the winning Qwen... |
external-community:cmovsdrrh00bup101f3th6aqx | external-community | cmovsdrrh00bup101f3th6aqx | 2026-08-10T02:44:31Z | 2026-05-07T17:55:09.390Z | 1 | Qwen/Qwen3.6-35B-A3B | main | Qwen3.6-35B-A3B | null | Qwen | 36 | 3 | true | llama.cpp | b9033 | cpu | UD-Q3_K_M | Intel R Core TM Ultra 7 155H | CPU_ONLY | null | 1 | null | null | null | Intel(R) Core(TM) Ultra 7 155H | null | Windows 11 Home (build 26200) | 256 | 128 | 2,048 | 1 | null | null | 13.36 | 35.79 | null | null | null | null | null | [] | null | null | null | null | null | null | false | false | dpenev98 | false | Sequential single-benchmark run on a Lenovo ThinkPad E16 (21MA003QBM) with Intel Core Ultra 7 155H, 32 GB RAM, and Windows 11 Home build 26200. Fixed profile: p=256, n=128, r=3, t=16, b=256, ub=256, ctk=q8_0, ctv=f16, mmp=0. Final baseline runs were taken in Windows max-performance mode. Best overall backend for this Q... |
external-community:cmp04pw1700f5lo01843tbmjh | external-community | cmp04pw1700f5lo01843tbmjh | 2026-08-10T02:44:31Z | 2026-05-10T18:51:34.891Z | 1 | Qwen/Qwen3.6-27B | main | Qwen3.6-27B | null | Qwen | 28 | null | false | vllm | 0.20.1 | cuda | NVFP4 | RTX 5090 | DISCRETE_GPU | RTX 5090 | 1 | 32 | null | null | Ryzen 7 3700X | null | Ubuntu 24.04 | 17 | 1,024 | 243,040 | 1 | null | null | 55.4 | 7,457.1 | 56.3 | null | null | null | 28.3 | [] | null | 1 | null | fp8_e4m3 | null | true | false | false | slow4cyl | false | Qwen3.6-27B-NVFP4 via vLLM 0.20.1 in Docker (WSL2). Compressed-tensors quantization, FP8 KV cache, RTX 5090 32GB. Reasoning model — output via reasoning delta. Prefill: 512 tokens in 0.068s. Decode: 1024 tokens in 18.48s. |
external-community:cmp0ivytq00otlo01ltt7r5vh | external-community | cmp0ivytq00otlo01ltt7r5vh | 2026-08-10T02:44:31Z | 2026-05-11T01:28:13.070Z | 1 | Qwen/Qwen3.6-35B-A3B | main | Qwen3.6-35B-A3B | null | Qwen | 36 | 3 | true | vllm | 0.20.2 | cuda | NVFP4 | RTX 5090 | DISCRETE_GPU | RTX 5090 | 1 | 32 | null | null | Ryzen 7 3700X | null | Ubuntu 24.04 | 5 | 512 | 262,144 | 1 | null | 24 | 170.1 | 10,828.7 | 171.7 | null | null | null | 30.2 | [] | null | 1 | null | null | null | null | false | false | slow4cyl | false | RTX 5090, WSL2, vLLM 0.20.2 V1 engine, NVFP4 quantization, gpu-memory-utilization=0.94, chunked prefill + prefix caching enabled, cudagraph capture size 32, max-num-seqs 16, max-num-batched-tokens 2048. Temperature 0 for deterministic throughput. Model loaded from /models/Qwen3.6-35B-A3B-NVFP4 in Docker container with ... |
End of preview. Expand in Data Studio
llmfit Real-World LLM Inference Benchmarks
An evolving dataset of real-world LLM inference measurements across consumer, workstation, datacenter, and unified-memory hardware. It combines benchmarks from an external community source with community benchmarks contributed directly to llmfit.
The initial release contains 1,501 normalized observations:
- 1,010 unique external-community observations from the repository's 2026-08-10 snapshot.
- 491 llmfit-community observations from 61 validated submissions, including repository updates through 2026-08-19.
The external source reported 2,473 total measurements across 27 hardware presets at snapshot time, but the embedded llmfit cache retains at most 100 rows per preset. It contains 1,111 cached rows, of which 101 are exact duplicates appearing in overlapping hardware-preset buckets. This dataset publishes the 1,010 unique records; it does not represent the uncached records as available observations.
Sources and permission
Every row has a source value of external-community or llmfit-community, plus a stable source record identifier. The dataset maintainers received permission to redistribute the external benchmark data in this combined dataset; the source requested not to be named publicly. llmfit-community submissions are contributed through the llmfit repository and validated against its published schema.
This dataset uses license: other because it combines data from separately administered sources. The llmfit software repository is MIT licensed; that software license should not be interpreted as overriding rights attached to third-party model names, benchmark inputs, or source data.
Intended use
The dataset supports:
- Model-to-hardware fit analysis
- Throughput and latency estimation
- Estimate calibration against measured performance
- Runtime and quantization comparisons
- Hardware purchasing and deployment research
- Reproducibility studies across drivers and runtime versions
It is not a controlled leaderboard. Measurements come from heterogeneous machines, prompts, runtime configurations, software versions, thermals, and user environments.
Schema
The unified table preserves source provenance while normalizing common concepts:
- Identity:
benchmark_id,source,source_record_id,measured_at - Model:
model_id, revision, family, parameter counts, MoE metadata - Runtime: engine/provider, version, backend, quantization, engine options
- Hardware: accelerator, count, memory, CPU, OS, hardware class
- Workload: prompt/output tokens, context length, batch size, run count
- Results: output/prefill/total throughput, TTFT, duration, VRAM, and power
- Provenance: source snapshot, submitter name where already public, and verification status
Unavailable source fields are represented as null rather than inferred. Estimates and simulated results are not inserted as measured observations.
Updating
The export is deterministic. Run:
python export.py /path/to/llmfit data/benchmarks.parquet
The llmfit community directory is append-only. Stable benchmark_id values make later exports suitable for deduplication and longitudinal use.
Limitations
- External-community data is a capped snapshot rather than a complete historical export.
- llmfit-community submissions currently contain aggregated runs rather than one row per raw run.
- Many records omit driver, runtime version, prompt, context, power, or peak-memory details.
- User-submitted measurements may be noisy or incorrectly identified.
- Results should not be treated as direct model-quality comparisons.
Citation
If this dataset is useful, cite the dataset revision you used and link to llmfit.
- Downloads last month
- -