feanors/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-AWQ-INT4

Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled — AWQ-INT4 (W4A16, multimodal)

INT4 quantization (AutoAWQ GEMM layout, W4A16, group_size 128, asymmetric) of lordx64/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled, a 35B-parameter / 3B-active Qwen3.5-MoE multimodal model that has been reasoning-distilled from Claude 4.7 Opus.

This artifact preserves the full multimodal stack (vision tower, linear-attention layers, shared experts, layer 0, MTP head, router gates) at fp16/bf16 and only quantizes the routed expert MLPs (gate_proj / up_proj / down_proj per expert × 256 experts × ~39 routed-MoE layers). The on-disk layout mirrors QuantTrio/Qwen3.6-35B-A3B-AWQ exactly so vLLM auto-detects and runs it with the same moe_wna16 / AWQ kernel path. If you don't need multimodal and want maximum quality, use the sibling FP8 build — feanorscode/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-FP8-Dynamic.

Method note — read this before citing as "AWQ". This build does data-free RTN quantization packed in the AutoAWQ GEMM format. It uses the AWQ on-disk layout and the AWQ-aware vLLM kernel path, but does not run AWQ's defining activation-aware salience pass over a calibration corpus. See Quantization recipe.

Status — smoke build, full eval done (2026-04-27). GSM8K 0.9386 (strict), MMLU 0.8068, ARC-C 0.5648 (acc) on the full benchmark sets. Strong enough that this checkpoint already looks production-usable; a calibrated re-run is on the roadmap.

---

Source code & reproduction

Quantized by FeanorsCode (github.com/FeanorsCodeSL) as part of the public DGX-Spark quantization framework: github.com/FeanorsCodeSL/dgx-spark-quantization.

The framework includes the eval driver, vLLM container launcher, the recipes for both the FP8 and AWQ-INT4 builds of this base, the full eval results (GSM8K + MMLU + ARC-Challenge), and the publishing walkthrough used to produce this artifact. If you want a vLLM-loadable quantization of a different bf16 / GGUF model, fork it.

• Per-scheme reference: docs/schemes/awq-gemm.md
• Recipe used for this build: runs/qwen3.6-35b-distill/recipes/awq_gemm.py
• Three-way bf16 / FP8 / AWQ comparison report: runs/qwen3.6-35b-distill/REPORT.md

---

TL;DR

	bf16 (base)	AWQ-INT4 (this build)
Bits / weight (routed experts)	16	4
Activation precision	16	16 (W4A16)
Calibration	n/a	none (data-free RTN)
Vision tower preserved	yes	yes (kept fp16)
MTP head preserved	yes	yes (kept fp16)
Layer 0 quantized?	n/a	no (kept fully fp16)
Disk size	~67 GiB	~24 GiB
Format	safetensors / bf16	AutoAWQ GEMM (moe_wna16 consumer)
GSM8K (1,319, 5-shot CoT, strict)	0.9447	0.9386
MMLU overall (14,042, 5-shot raw MC)	0.8341	0.8068
ARC-Challenge (1,172, raw MC, acc)	0.5478	0.5648
Δ MMLU vs bf16	—	−2.73 pp (~9 σ — real)
Δ GSM8K strict vs bf16	—	−0.61 pp (~1 σ)

Same eval harness (lm-evaluation-harness 0.4.11), same prompts, same temperature=0, same KV-cache dtype.

---

Quick numbers

benchmark	metric	score	n
GSM8K (5-shot CoT, chat-templated)	exact_match strict	0.9386 ± 0.0066	1,319 (full)
GSM8K (5-shot CoT, chat-templated)	exact_match flexible	0.9416 ± 0.0065	1,319 (full)
MMLU overall (5-shot, raw multiple-choice loglikelihood)	acc	0.8068 ± 0.0032	14,042 (full, 57 subtasks)
ARC-Challenge (raw multiple-choice loglikelihood)	acc	0.5648 ± 0.0145	1,172 (full)
ARC-Challenge (raw multiple-choice loglikelihood)	acc_norm	0.5606 ± 0.0145	1,172 (full)

Full eval, no subsetting. See Evaluation vs bf16 for methodology and the bf16 deltas.

---

Model summary

Base model	lordx64/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled
Underlying base	Qwen/Qwen3.6-35B-A3B (Apache-2.0)
Architecture	Qwen3_5MoeForConditionalGeneration (multimodal: text + vision)
Total params	~35 B
Active params / token	~3 B (top-8 of 256 experts)
Hidden size	2,048
MoE intermediate	512
Layers	40 (12 full-attention + 28 linear-attention, repeating)
Experts	256 routed + 1 shared, top-8 routing
Vocab	248,320
Max position	262,144 (RoPE θ = 10,000,000)
Vision	27-block ViT, hidden 1,152, patch 16, m-RoPE — kept fp16
MTP head	1 layer — kept fp16
Quantization	AWQ-GEMM, W4A16 asymmetric, group_size 128, zero_point true
Quantized scope	routed expert MLPs only
Kept fp16/bf16	visual.*, linear_attn.*, self_attn.*, shared_expert.*, mlp.gate (router), model.layers.0.*, mtp.*, lm_head
On-disk size	~24 GiB (vs. ~67 GiB bf16 source, ~35 GiB FP8 sibling)

---

Quantization recipe

Method (precise wording)

Custom shard-streaming data-free round-to-nearest (RTN) quantizer, packed in the AutoAWQ GEMM layout so vLLM's existing AWQ / moe_wna16 kernels consume it unchanged.

This is not the activation-aware salience pass that defines AWQ in the Lin et al. 2023 paper. We use:

• AWQ's on-disk layout (qweight / qzeros / scales with the canonical 8-nibble pack permutation [0, 4, 1, 5, 2, 6, 3, 7]).
• AWQ's runtime kernel path in vLLM (auto-detected from quantization_config.quant_method = "awq").
• Plain RTN for the math — per-group asymmetric min/max, no calibration corpus, no salience scaling.

A separate calibrated AWQ build (real activation-aware pass over a small reasoning + chat corpus) is on the roadmap.

Math (per in-scope [out, in] weight, per expert sub-tensor)

Per-group asymmetric quantization along in_features with group_size = 128:

scale  = (max(W_g) - min(W_g)) / 15        # round-trip via fp16
zp     = round(-min(W_g) / scale)          # int, clamped to [0, 15]
q      = clip(round(W / scale) + zp, 0, 15)

Then 8 nibbles along out_features are packed into one int32 using AWQ's canonical lane permutation [0, 4, 1, 5, 2, 6, 3, 7], producing the exact qweight / qzeros / scales tensor layout vLLM's convert_awq_tensor (moe_wna16.py) consumes.

The fused 3-D Qwen3.5-MoE expert tensors are unfused on the way out:

• experts.gate_up_proj ([256, 2·512, 2048]) is split into 256 × gate_proj + 256 × up_proj ([512, 2048] each).
• experts.down_proj ([256, 2048, 512]) is split into 256 × down_proj per layer.

Settings

key	value
quant_method	awq
version	gemm
bits	4
group_size	128
zero_point	true
asymmetric	yes (per-group min/max)
symmetric	no
pack order	[0, 4, 1, 5, 2, 6, 3, 7] (AWQ canonical)
calibration data	none (data-free RTN)
calibration samples	0
modules_to_not_convert	["visual", "linear_attn", "self_attn", "shared_expert", "mlp.gate", "model.layers.0.", "mtp"]
activation precision	fp16 (W4A16)
storage dtype	int32 (packed nibbles) + float16 (scales)
MoE expert layout	unfused → per-expert gate_proj / up_proj / down_proj

Why these specific exclusions

module class	reason for keeping fp16
visual.*	Multimodal vision tower; vision quality degrades sharply under naive INT4.
linear_attn.*	Mamba-style state — INT4 destabilizes the recurrence.
self_attn.*	Q/K/V/O are the largest single bottleneck for downstream quality on attention layers; cheap to keep at fp16 because there are only 12 full-attn layers.
shared_expert.*	Always-active dense MLP path — kept full precision for stable baseline routing.
mlp.gate	Router gate (the small [hidden, num_experts] projection). Quantizing it would corrupt routing decisions.
model.layers.0.*	Earliest layer is the most sensitive in MoE models — by convention kept dense.
mtp.*	Speculative-decoding head; small and quality-critical for accept rates.
lm_head	Output embedding — standard practice.

This list matches QuantTrio's reference build of the same architecture so vLLM's MoE-WNA16 / AWQ kernels recognize the layout without tweaks.

Hardware & runtime

• Quantizer runtime: NVIDIA DGX Spark (single node, GB10 / SM121a, 128 GiB unified memory), CPU-only quantization path — no GPU needed because the math is pure RTN over per-group min/max.
• Source streaming: one bf16 safetensors shard at a time (~6 source shards, each ≤ 5 GiB), so peak host RAM stayed well under 24 GiB.
• Output: 6 shards × ~4 GiB each, total ~24 GiB.

---

Inference (vLLM)

vLLM auto-detects quantization_config.quant_method = "awq" from config.json. Tested with the official vllm/vllm-openai container:

docker run --gpus all --ipc=host --rm \
  -p 8000:8000 \
  vllm/vllm-openai:latest \
  --model feanorscode/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-AWQ-INT4 \
  --served-model-name qwen3.6-awq \
  --quantization awq \
  --dtype float16 \
  --gpu-memory-utilization 0.8 \
  --max-model-len 4096 \
  --trust-remote-code

Then hit http://localhost:8000/v1/chat/completions with the OpenAI chat format. (/v1/completions works for raw text too, but multimodal inputs require the chat endpoint.)

Notes:

• --quantization awq selects vLLM's AWQ / moe_wna16 kernel path.
• --dtype float16 matches the saved scale dtype.
• max-model-len is capped at 4096 in our smoke config for memory headroom on consumer-grade GPUs; the underlying model supports 262 K.

---

Evaluation vs bf16 baseline

Full eval (2026-04-27, total wall-clock 1 h 33 min) via lm-evaluation-harness 0.4.11 against the live vLLM endpoint (local-completions for raw multiple-choice tasks, local-chat-completions for generation tasks that need the chat template). vLLM was served from the community pre-built DGX-Spark image ghcr.io/spark-arena/dgx-vllm-eugr-nightly-tf5:20260415, built from eugr/spark-vllm-docker (MIT) — no upstream vLLM bare-metal build is yet stable on aarch64 / SM121a. Same harness, same prompts, same fp8_e4m3 KV-cache dtype as the bf16 baseline run.

Setup

harness	lm-evaluation-harness==0.4.11 (+ [api] extras: tenacity)
backend	vLLM --quantization awq --dtype float16
max_length	4,096
num_concurrent	4
sampling (gen tasks)	temperature=0, top_p=1, max_gen_toks=1024
tokenizer	huggingface backend pointed at this artifact's local dir

Headline numbers vs bf16

benchmark	metric	bf16	AWQ-INT4	Δ (AWQ − bf16)
GSM8K (5-shot CoT, chat-templated)	exact_match strict	0.9447	0.9386	−0.61 pp (~1 σ)
GSM8K (5-shot CoT, chat-templated)	exact_match flexible	0.9447	0.9416	−0.31 pp (within σ)
MMLU overall (5-shot, raw MC)	acc	0.8341	0.8068	−2.73 pp (~9 σ — real)
MMLU humanities	acc	0.7751	0.7458	−2.93 pp
MMLU social sciences	acc	0.9074	0.8902	−1.72 pp
MMLU STEM	acc	0.8208	0.8059	−1.49 pp
MMLU other	acc	0.8642	0.8175	−4.67 pp
ARC-Challenge (raw MC)	acc	0.5478	0.5648	+1.70 pp (within stderr)
ARC-Challenge (raw MC)	acc_norm	0.5648	0.5606	−0.42 pp (within stderr)

Reading the deltas. The headline cost of INT4 routed-experts is on MMLU — −2.7 pp overall, with the largest single hit on the "other" cluster (general-knowledge subtasks, −4.67 pp), consistent with INT4 routed-experts losing some long-tail-knowledge precision. STEM and social sciences hold tighter (−1.5 pp). GSM8K barely moves — reasoning chains stay coherent at INT4 weights / fp16 activations. ARC-Challenge is a tie (both deltas inside ±1.45 pp stderr).

MMLU best / worst subtasks (AWQ build)

Best: high_school_government_and_politics 0.974, high_school_microeconomics 0.958, conceptual_physics 0.945.

Weakest: virology 0.548, global_facts 0.550, moral_scenarios 0.590.

(Several weak subtasks are also weak on the unquantized base — they don't necessarily reflect quantization damage.)

Quantization-cost benchmark in plain terms

You're paying ~2.7 pp on MMLU to get:

• Disk size from 67 GiB → **24 GiB** (~2.8× compression).
• The full multimodal stack preserved (vision + MTP).
• The fastest decode path of the three builds (1 h 33 min for the eval battery vs FP8's 1 h 59 min vs bf16's 2 h 45 min on identical hardware).

If MMLU points matter more than disk and multimodal, ship the FP8 sibling instead.

Eval limitations

• Single seed. No variance over seeds.
• No vision evals. The vision tower is unquantized so multimodal performance should match the base, but this hasn't been confirmed end-to-end on standard multimodal benchmarks.
• No long-context evals. Truncated to 4 K context to fit the smoke serving config; the model itself supports 262 K.

---

Limitations

• Data-free RTN, not calibrated AWQ. This build does not run an activation-aware salience pass over a calibration corpus. The strong benchmark numbers indicate the routed-expert weight distribution is benign enough that naive RTN is already close to optimal, but on out-of-distribution / low-resource inputs you may see more degradation than a calibrated AWQ build would.
• Vision/language drift unmeasured. The vision tower stays at fp16, but interactions between fp16 vision features and INT4 expert MLPs in downstream layers haven't been benchmarked on multimodal tasks.
• Smoke serving config ≠ production config. Long-context throughput, batch-size sweeps, and m-RoPE behavior at full 262 K weren't part of this round.
• Reasoning-distilled, not RLHF'd. Inherits all the safety/style characteristics of the base distill, including any quirks of distilling Claude 4.7 Opus reasoning traces into a 3 B-active MoE.
• Anthropic usage-policy obligation propagates. See License & usage.

---

Roadmap

1. Full-set re-runs. Done (2026-04-27).
2. BF16 baseline diff. Done (2026-04-27): MMLU −2.73 pp, GSM8K strict −0.61 pp, ARC-C tie. See table above.
3. Calibrated AWQ build. Re-quantize with an activation-aware salience pass (small calibration corpus, e.g. ~256 sequences from a mix of reasoning and chat data) — likely to recover the ~2.7 pp MMLU gap.
4. Code coverage. HumanEval in a sandboxed container.
5. Long-context probe. RULER / NIAH at ≥ 32 K to confirm the linear- attention layers tolerate the W4A16 routed experts in long-context settings.

---

Files

config.json                   # arch + quantization_config (matches QuantTrio AWQ)
chat_template.jinja           # passed through from base
tokenizer.json                # passed through from base
tokenizer_config.json         # passed through from base
processor_config.json         # passed through from base (multimodal)
model.safetensors.index.json
model-00001-of-00006.safetensors
model-00002-of-00006.safetensors
model-00003-of-00006.safetensors
model-00004-of-00006.safetensors
model-00005-of-00006.safetensors
model-00006-of-00006.safetensors
README.md                     # this file

---

License & usage

Model weights: Apache-2.0, inherited from the base model lordx64/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled, which in turn inherits from Qwen/Qwen3.6-35B-A3B (also Apache-2.0).

Quantization code: Apache-2.0 — published by FeanorsCode at FeanorsCodeSL/dgx-spark-quantization.

Distillation-data provenance — read this before deploying. The base model's training data was generated using Anthropic's Claude Opus 4.7 via API. lordx64's model card explicitly states that "downstream users should confirm compliance with Anthropic's usage policies for their specific use case." That obligation propagates to this quantized derivative. If you intend to use this model in production — particularly for any commercial or model-training purpose — verify that your use case is compatible with Anthropic's usage policy as of your deployment date.

---

Citation (this build)

@misc{qwen36_distill_claude47_awq_int4_2026,
  title  = {Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled --- AWQ-INT4 (W4A16, multimodal, RTN-packed-AWQ-format)},
  author = {feanors (FeanorsCode)},
  year   = {2026},
  url    = {https://huggingface.co/feanorscode/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-AWQ-INT4},
  note   = {Data-free RTN INT4 quantization of lordx64/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled, packed in AutoAWQ GEMM format for vLLM compatibility}
}

If you cite this build, please also cite the AWQ paper (above), the base model, and Qwen3.5-MoE.

---

Credits

This build stands on a stack of open-source work. Listed in the order each piece appears in the pipeline.

Model lineage

• Qwen team — the Qwen3.5-MoE architecture and the Qwen/Qwen3.6-35B-A3B base (Apache-2.0).
• lordx64 — the Claude-4.7-Opus reasoning-distilled base this build quantizes (Apache-2.0).
• Anthropic — the reasoning-trace data source (Claude Opus 4.7 via API, used by lordx64 to train the distill). See License & usage for the propagated policy obligation.

Quantization (AWQ-INT4 / W4A16)

• AWQ method — Lin, Tang, Tang, Yang, Chen, Wang, Xiao, Dang, Gan, Han, "AWQ: Activation-aware Weight Quantization for LLM Compression and Acceleration", arXiv:2306.00978,

  2023. This build uses the AWQ format and kernel path but does not run AWQ's defining activation-aware salience pass — the math is plain RTN. See Method.

  @article{Lin2023AWQ,
    title  = {AWQ: Activation-aware Weight Quantization for LLM Compression and Acceleration},
    author = {Lin, Ji and Tang, Jiaming and Tang, Haotian and Yang, Shang and
              Chen, Wei-Ming and Wang, Wei-Chen and Xiao, Guangxuan and
              Dang, Xingyu and Gan, Chuang and Han, Song},
    journal= {arXiv preprint arXiv:2306.00978},
    year   = {2023}
  }
  

• AutoAWQ — MIT, by casper-hansen. Source of the on-disk packing layout (qweight / qzeros / scales, canonical 8-nibble pack permutation [0, 4, 1, 5, 2, 6, 3, 7]). The repository was archived in May 2025; vLLM's moe_wna16 kernels still consume the format directly.

• QuantTrio/Qwen3.6-35B-A3B-AWQ — layout reference. This build's modules_to_not_convert list and per-expert unfusing scheme match QuantTrio's reference exactly so vLLM consumes both checkpoints through the same code path.

Serving & evaluation infrastructure

• vLLM — Apache-2.0, vllm-project. Serving stack and the moe_wna16 AWQ kernel path that loads this checkpoint.
• eugr/spark-vllm-docker — MIT, by eugr. The community pre-built DGX-Spark vLLM container image (ghcr.io/spark-arena/dgx-vllm-eugr-nightly-tf5:20260415) used to serve this build during evaluation.
• lm-evaluation-harness — MIT, by EleutherAI. Eval framework and benchmark implementations (GSM8K, MMLU, ARC-Challenge) used for the bf16 ↔ AWQ comparison.
• Hugging Face — transformers / safetensors / the Hub itself (Apache-2.0).
• PyTorch — BSD-3-Clause. Tensor primitives used by the streaming quantizer.

Hardware

• NVIDIA DGX Spark (GB10 / SM121a, 128 GiB unified memory) for both the quantization and the evaluation runs. This repository is not affiliated with NVIDIA.

No method patents are involved. All cited works are open-source under permissive licenses (Apache-2.0, MIT, BSD-3-Clause).