Oct 24 2 days ago 11:08 onw 3.3.0: LFM2.5-24B-A2B (converted on this PC), faster MoE prompts Oct 20 6 days ago 13:02 onw 3.2.1: faster MoE answers (experts kept in their slots, bound attention for MoE, LM head compiled with profiling on), the Qwen3.6-35B-A3B NF4 build in the catalog 02:16 onw 3.2.0: Ternary Bonsai 2 4B / 9B 64-token graphs (one copy of the 2-bit weights), parallel compile in processes, faster kit compiler, paged graphs released under fused ones, vision encoder in the background Oct 19 7 days ago 09:09 onw 3.1.0: the Qwen family on fused graphs + MTP (setting, on by default), background preload of fused variants up to 4K, model list search / sort / filters 02:45 onw 3.0.0: onw NPU kit (Windows and Linux) - weightless graphs, shared weights, compiler fixes; fused graphs on by default from Lunar Lake on, up to 32K positions Oct 18 8 days ago 10:48 onw 2.1.0: NF4 Granite 4.2 3B / 8B and llm-jp-4.1-8B (Lunar Lake and newer); wired dense decode, bound attention, PLD in the 2-token row; fused graphs experimental (ONW_FUSED=1) Oct 17 9 days ago 22:38 onw 2.0.3: NPU 4000+ dense models answer on 2-token graphs (Granite 4.2 / llm-jp-4.1-8B first load on Panther Lake 2 h 45 min -> 23 min), known issues updated 15:02 README: known issues (onw 2.0.2) - slow first load of dense models on Panther Lake (NPU compiler 8.4 write-out), MTP decode on small Panther Lake PCs, NF4 faster on Panther Lake, Qwen3.6 speed-up needs 24 GB 08:02 onw 2.0.2: dense 64-token prompt blocks on all NPUs again - NPUW_DQ_FULL=NO on NPU 5000+ (the cause of the Panther Lake errors); Panther Lake Qwen3.5-9B 44 -> ~157 tok/s Oct 16 10 days ago 20:38 onw 2.0.1: dense 64-token prompt blocks only on the NPU 3720 (they gave wrong output on Panther Lake NPU 5010); NPU 4000+ back to the 1.4 block sizes 20:29 onw 2.0.0: Qwen3.6 MoE prompts ~4-6x (grouped experts), dense prompts ~2-2.5x (64-token blocks), faster MoE decode 05:02 onw 1.4.13: Ternary Bonsai 2 9B / 4B (2-bit, images, Lunar Lake and newer); onw convert reads their packs Oct 11 15 days ago 04:16 onw 1.4.9: Ternary Bonsai 2 27B (2-bit weights on NPU 4000+, Hadamard in-graph, onw convert reads Prism ML MLX packs); chat max tokens default to the context; settings marked that recompile; driver/OpenVINO update notice Oct 6 20 days ago 11:49 onw 1.4.5: first compile ~3 min shorter (4/64-token DeltaNet parts go straight to one-graph NPUW), time-left estimate while compiling; README: Panther Lake compile-time note with a Lunar Lake comparison 01:23 README: the power chart at the top too (before Install) 01:21 README: power chart - NPU highlighted, grouped by draft-token setting, average temperature 01:13 README: power efficiency with onw 1.4.0 (NF4 + MTP) - without / with draft tokens on NPU, iGPU, CPU; chart Oct 5 21 days ago 23:37 onw 1.4.0: MTP draft tokens (~1.5x answers on Lunar Lake, NF4 Qwen3.5 family), 64-token prompt blocks on Lunar Lake, less host work per step, fix all-zero logit rows 05:38 README: add power efficiency and temperature section (Qwen3.5-9B, Lunar Lake) Oct 4 22 days ago 13:08 onw 1.3.0: Anthropic Messages API and OpenAI Responses API, thinking levels with budgets, per-model defaults, chat copy/source/regenerate 11:21 onw 1.2.0: llm-jp-4.1-32B-A3B-thinking (qwen3_moe) Oct 3 23 days ago 21:46 onw 1.1.1: Qwen3.5-9B NF4, recommended sampling in the chat 15:36 onw 1.1.0: NF4 models for Lunar Lake and newer (Ornith-1.5-9B NF4) 13:14 onw 1.0.1: user guide and new screenshots, chat formatting, Gemma repetition hint 10:05 onw 0.15.1: memory-saving mode loads one block size at startup, strict memory-saving mode 08:07 onw 0.15.0: chat history, reasoning_content always + reasoning_tokens, thinking info and default, llm-jp INT8 05:34 onw 0.14.2: memory-saving mode for every dense model without shared weights, Gemma 4 thinking split, stale-tray warning, LM head via npuw-none first 04:10 onw 0.14.1: memory-saving mode swaps the 16-token and 1-token graphs (one set in memory at a time) 03:13 onw 0.14.0: memory-saving mode (16-token graphs only) for many-part models; no auto-reload after updates when auto-load is off Oct 2 24 days ago 23:19 onw 0.13.0: NeoHorse-1-4B, Granite 4.2 3B / 8B, llm-jp-4.1-8b-thinking; granite / llama converter; Panther Lake listed Oct 1 25 days ago 09:52 onw 0.12.0: Ornith-1.5-9B, earlier turns' reasoning_content to the template, --deltanet-fmt q8g128 Sep 29 27 days ago 14:42 onw 0.11.1: installs again (OpenVINO pre-release index), code review fixes, gpt-oss INT8 embedding 12:32 onw 0.11.0: OpenAI gpt-oss-20b (gpt_oss: MoE, attention sinks, YaRN, harmony), INT8 code overflow fixes Sep 28 28 days ago 23:51 onw 0.10.1: experimental NPUW settings (MoE, FUNCALL_ASYNC, weightless cache); less CPU in long chats 18:37 onw 0.10.0: long contexts with a paged INT8 KV cache (32K default, up to the model max); Gemma RoPE fix; Ubuntu launcher fix 15:07 onw 0.9.2: tool calls (function calling) in the OpenAI-compatible API 03:43 onw 0.9.0: LFM2.5 support (convert it yourself), LFM models out of the list, torch-free convert 01:23 README reorganized (install first); technical notes moved to TECHNICAL.md Sep 27 29 days ago 11:29 onw 0.8.4: English / Simplified Chinese UI, new icon, README banner 11:03 README: screenshots of the onw window 09:43 v0.8.2: predicted prefetch while decoding off by default (CPU cost on 8-core Lunar Lake); prompt prefetch stays 08:49 v0.8.0: faster SSD streaming without more memory - prompt prefetch into cache slots, predicted next-layer expert prefetch while decoding (helper thread, separate read queue), single-copy reads Sep 26 30 days ago 16:22 v0.7.0: 20-60% less memory (own blob store, vision loaded on demand and freed, host tables mapped, direct-I/O expert load); REAP streams too; first-on-Intel-NPU SSD streaming note 15:19 v0.6.1: layer-major prompt processing when experts stream from the SSD (each layer's experts read once per prompt) 15:13 v0.6.0: stream MoE experts from the SSD (LRU cache, direct I/O) when the model does not fit; expert memory setting Sep 25 15:10 README: to our knowledge the first public all-on-NPU VLM / Gated DeltaNet / MoE (survey Sept 2026) 11:59 v0.4.8: NPUW fixes (no SLICE_OUT, bool constants), NaN check, ONW_SIZES, Linux driver 1.38 note, diagnostic tools 07:04 v0.4.7: no NPUW for MoE models, bounded hold on broken characters, clear-cache button, save button on top 06:22 v0.4.5: NPUW weight sharing (auto on NPU 4000+, experimental), taller list rows on Ubuntu 05:44 v0.4.4: load/unload UI with model locations, single instance, background polling, per-PC memory estimate 
