README.md

metadata

base_model: nex-agi/Nex-N2-mini
license: apache-2.0
library_name: gguf
tags:
  - gguf
  - rocmfp4
  - qwen3.5
  - nex-n2
  - coder
  - agentic
  - moe
  - imatrix
  - strix-halo
  - amd
  - rocm
  - vulkan
language:
  - en
base_model_relation: quantized

PLUNDERSTRUCK // ROCmFP4 QUANTIZED MODEL // STRIX HALO · gfx1151

███████▆▄▁     ██████▎▐██████████████▖▀██████▙      ▟██████▘
██████████▆▃▁  ██████▎▐███████████████▙▝▜██████▖  ▗██████▛  
█████████████▆ ██████▎▐████▀▀▀▀▀▀▀▀▀▀▀▀▀ ▀██████▙▟██████▘   
██████▏▀▜█████ ██████▎▐████▃▃▃▃▃▃▃▃▃▃▃▃▃▖ ▔▜██████▘ ▀█▛     
██████▏▗▂▔▀▜██ ██████▎▐█████████████████▍   ▝███▛▁▟▙        
██████▏▐█▇▅▂▔▀ ██████▎▐█████████████████▍    ▔▀▘▃████▖      
██████▏▐████▇▄▂██████▎▐████▀▀▀▀▀▀▀▀▀▀▀▀▀  ▗▟█▖▁▟██████▙▁    
██████▏▝█████████████▎▐████▅▅▅▅▅▅▅▅▅▅▅▅▅ ▄██████▛▜██████▖   
██████▏  ▝▜██████████▎▐███████████████▛▗▟██████▘  ▝██████▙▁ 
██████▏     ▀▜███████▎▐██████████████▘▄██████▛      ▜██████▖

NEX-N2-MINI

4-BIT ROCmFP4 · CODE-WEIGHTED IMATRIX · HIGH-SPARSITY MoE (3B ACTIVE) · AGENTIC CODER · SINGLE AMD APU

FORMAT ROCmFP4 4-BIT	PRECISION ~4.5 BPW	SIZE 18.4 GB	CONTEXT 131 K
ARCH qwen35moe	PARAMS 35B / 3B ACTIVE	BACKEND VULKAN0	LICENSE APACHE-2.0

⚠ REQUIRES THE ROCmFP4 FORK
The custom q4_0_rocmfp4 / q4_0_rocmfp4_fast tensor types will not load in stock llama.cpp, LM Studio, or Ollama. Build/run with charlie12345/ROCmFPX · branch mtp-rocmfp4-strix.

NOTE // Ignore HuggingFace's auto-detected "F16" badge — its parser can't read ROCmFP4 and mislabels by the f16 embeddings. These are ~4.5 bpw 4-bit files; pick by filename.

01 · FILES

File	Size	Output head	Pick if
…-STRIX-embF16-imatrix-headQ6.gguf ★	18.4 GB	Q6_K	the one build — best speed/quality balance: f16 embeddings + Q6 output head on the fast single-scale body

One file — the best speed/quality balance in ROCmFP4 for Strix Halo. It keeps the two quality levers that are actually felt — genuine f16 token embeddings (from BF16) and a Q6_K output head — on the fast single-scale q4_0_rocmfp4_fast body + the code-weighted imatrix (see §04). Not the leanest-fastest possible (a 4-bit output head squeezes out a few more tok/s, at a fidelity cost), and not the most faithful possible (see the base-model fidelity link in §04) — it's the point where speed and quality meet best. The Qwen (ChatML) chat template is baked into the GGUF — just pass --jinja.

02 · QUICK START

Run from the folder holding the .gguf:

env HSA_OVERRIDE_GFX_VERSION=11.5.1 GGML_HIP_ENABLE_UNIFIED_MEMORY=1 \
llama-server \
  -m Nex-N2-mini-ROCmFP4-STRIX-embF16-imatrix-headQ6.gguf \
  --alias nex-n2-mini \
  --host 0.0.0.0 \
  --port 8080 \
  -dev Vulkan0 \
  -ngl 999 \
  -fa on \
  -c 131072 \
  -b 2048 \
  -ub 256 \
  -t 16 \
  -tb 16 \
  -ctk f16 \
  -ctv f16 \
  -cpent 256 \
  -ctxcp 32 \
  --cache-reuse 256 \
  --cache-ram 65536 \
  --temp 0.6 \
  --top-p 0.95 \
  --top-k 20 \
  --min-p 0.0 \
  --jinja \
  --parallel 1 \
  --metrics \
  --no-mmap

NOTE // No --spec-* / --spec-type draft-mtp flags — Nex-N2-mini ships without an MTP head (non-speculative). At ~72 t/s it doesn't need speculative decoding to be quick.

Flag	Function
HSA_OVERRIDE_GFX_VERSION=11.5.1	treat the APU as gfx1151 (Strix Halo)
GGML_HIP_ENABLE_UNIFIED_MEMORY=1	allow use of the full 128 GB unified memory
-dev Vulkan0	run on Vulkan — fastest backend for ROCmFP4 on Strix Halo
-ngl 999 · -fa on	offload all layers · flash attention
-c 131072	context length (128K)
-b 2048 · -ub 256 · -t/-tb 16	prefill batch / micro-batch · CPU threads
-ctk f16 · -ctv f16	f16 KV cache — how we run it; drop to q8_0/q4_0 to use less memory
-cpent · -ctxcp · --cache-reuse · --cache-ram 65536	cross-turn KV checkpointing + 64 GB resident reuse cache
--temp 0.6 --top-p 0.95 --top-k 20 --min-p 0.0	base-model recommended sampling
--jinja --parallel 1 --metrics --no-mmap	apply baked ChatML template · single slot · metrics · weights in RAM

03 · AGENTIC CODING / TOOLS

Nex-N2-mini is an agentic / "thinking" coder — agentic tool-use trained. To get native tool calls, your client must use the qwen3_coder tool-call parser. Without it the model tends to narrate code instead of emitting structured tool calls.

CHAT TEMPLATE	Qwen (ChatML) — baked into the GGUF; pass --jinja
TOOL-CALL PARSER	qwen3_coder — set in your client/runtime
SAMPLING	temp 0.6 · top-p 0.95 · top-k 20 (base-model recommended)

04 · PERFORMANCE & QUALITY

DECODE · short-context	~72 t/s (Vulkan / Ryzen AI Max+ 395)
SWE-BENCH VERIFIED · base model	74.4
ACTIVE PARAMS	3B of 35B (high-sparsity MoE)
QUANTIZATION	fast single-scale body + f16 embeddings + Q6 head + code-weighted imatrix

This is the best speed/quality balance in ROCmFP4 — by design, not the absolute fastest. It keeps the two quality levers that are actually felt — genuine f16 token embeddings and a Q6_K output head — on the fast single-scale q4_0_rocmfp4_fast body. A leaner 4-bit-output-head build is a few tok/s faster but degrades fidelity you'll notice; an all-dual-scale body buys a KL improvement that sits inside the measurement noise while costing decode speed. The fast body + f16 embeddings + Q6 head is the point where those meet best.

How we landed on this recipe. We ran the full body-kernel / head-precision / dual-scale sweep — KL divergence vs the BF16 reference plus llama-bench decode — on the dense Qwen3.6-27B sibling, where the same q4_0_rocmfp4 levers apply. The frontier there was unambiguous: the all-dual-scale body and selective higher-precision tensors both traded decode speed for a KL gain inside the noise, so the fast body + f16 embeddings + Q6 head won the balance. We carry that conclusion to this MoE rather than re-running the whole sweep per model — see the 27B sweep for the numbers and the format-limit reasoning. (Directional internal measurements — reproduce before citing.)

WANT MAXIMUM FIDELITY INSTEAD OF SPEED? Grab a Q6_K / Q8_0 GGUF of the base from nex-agi/Nex-N2-mini — those higher-bit GGUFs run on this same fork. We optimize for throughput in ROCmFP4; if you want the last bit of fidelity over speed, a higher-bit quant of the base is the one to grab.

The imatrix — code-weighted, and measured (it helps here). Quantized with an importance matrix from a code-weighted calibration mix (~2.6:1 code:general — eaddario code + Kalomaze groups_merged via froggeric/imatrix). Measured by KL-divergence + perplexity vs the true BF16 on a held-out code slice (disjoint from calibration):

Metric (vs BF16, held-out code)	No-imatrix	Imatrix	Change
Perplexity	4.076	4.013	−1.5% (recovers >½ the 4-bit loss; ~3.3σ)
Median KLD	0.0184	0.0159	−13%
RMS Δp	8.57%	8.00%	−7%
Same top token as BF16	88.97%	89.44%	+0.5 pp

For this model the imatrix is a clean win — better on every metric, including perplexity. (It's model-dependent — on the dense Qwopus-Coder the same recipe worsened code-PPL, so we shipped that one without imatrix. Always measure.)

05 · BUILD (REPRODUCIBLE)

# code-weighted imatrix on the BF16 (single pass)
llama-imatrix -m Nex-N2-mini-bf16.gguf -f code-weighted-calib.txt -o nexn2.imatrix -c 512 -ngl 999

# quant -> ROCmFP4 with the imatrix + genuine f16 embeddings
llama-quantize --token-embedding-type f16 --imatrix nexn2.imatrix \
  Nex-N2-mini-bf16.gguf \
  Nex-N2-mini-ROCmFP4-STRIX-embF16-imatrix.gguf  Q4_0_ROCMFP4_STRIX

# THE ONE BUILD (★): add the Q6_K output head on the fast single-scale body — best speed/quality balance (§04)
llama-quantize --token-embedding-type f16 --output-tensor-type q6_K --imatrix nexn2.imatrix \
  Nex-N2-mini-bf16.gguf \
  Nex-N2-mini-ROCmFP4-STRIX-embF16-imatrix-headQ6.gguf  Q4_0_ROCMFP4_STRIX

Experimental research build for AMD Strix Halo — hardware/driver/prompt-sensitive, may not reproduce elsewhere. Not native FP4 tensor-core execution.

06 · LINEAGE & CREDITS

BASE MODEL	nex-agi/Nex-N2-mini (Apache-2.0) · Qwen3.5-35B-A3B lineage (35B total / 3B active MoE)
CALIBRATION	eaddario/imatrix-calibration (code) + Kalomaze groups_merged via froggeric/imatrix (general)
FORMAT + RUNTIME	charlie12345/ROCmFPX (based on llama.cpp, MIT)

Derivative quantization — verify the base model's license before redistribution / use.