🔥 Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled

A reasoning SFT fine-tune of Qwen/Qwen3.6-35B-A3B on chain-of-thought (CoT) distillation mostly sourced from Claude Opus 4.6. The goal is to preserve Qwen3.6's strong agentic coding and reasoning base while nudging the model toward structured Claude Opus-style reasoning traces and more stable long-form problem solving.

The training path is text-only. The Qwen3.6 base architecture includes a vision encoder, but this fine-tuning run did not train on image or video examples.

• Developed by: @hesamation
• Base model: Qwen/Qwen3.6-35B-A3B
• License: apache-2.0

This fine-tuning run is inspired by Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled, including the notebook/training workflow style and Claude Opus reasoning-distillation direction.

Benchmark Results

The MMLU-Pro pass used 70 total questions per model: --limit 5 across 14 MMLU-Pro subjects. Treat this as a smoke/comparative check, not a release-quality full benchmark.

Benchmark	Harness	Samples per model	Setting	Metric	Base model	Fine-tuned merged model	Delta
MMLU-Pro overall	lm-evaluation-harness	70	--limit 5 across 14 subjects	exact_match, custom-extract	42.86%	75.71%	+32.85 pp

Base model: Qwen/Qwen3.6-35B-A3B. Fine-tuned model: hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled.

Community benchmarks welcome

To better understand this fine-tuned model's capabilities, I welcome independent benchmark results. If you run evaluations, please include the benchmark name, harness/script, sample count, decoding settings, and raw logs or result files when possible.

Share results by opening a PR/discussion or DMing @hesamation on X.

Base Qwen3.6 Highlights

This release delivers substantial upgrades, particularly in:

• Agentic Coding: the model now handles frontend workflows and repository-level reasoning with greater fluency and precision.
• Thinking Preservation: Qwen introduced a new option to retain reasoning context from historical messages, streamlining iterative development and reducing overhead.

For more details, please refer to the Qwen blog post Qwen3.6-35B-A3B.

Base Model Overview

• Type: Causal Language Model with Vision Encoder
• Training Stage: Pre-training & Post-training
• Language Model:
  • Number of Parameters: 35B in total and 3B activated
  • Hidden Dimension: 2048
  • Token Embedding: 248320 (Padded)
  • Number of Layers: 40
  • Hidden Layout: 10 x (3 x (Gated DeltaNet -> MoE) -> 1 x (Gated Attention -> MoE))
  • Gated DeltaNet:
    • Number of Linear Attention Heads: 32 for V and 16 for QK
    • Head Dimension: 128
  • Gated Attention:
    • Number of Attention Heads: 16 for Q and 2 for KV
    • Head Dimension: 256
    • Rotary Position Embedding Dimension: 64
  • Mixture Of Experts:
    • Number of Experts: 256
    • Number of Activated Experts: 8 Routed + 1 Shared
    • Expert Intermediate Dimension: 512
  • LM Output: 248320 (Padded)
  • MTP: trained with multi-steps
• Context Length: 262,144 natively and extensible up to 1,010,000 tokens.

Base Benchmark Results

The following table is from the upstream Qwen3.6-35B-A3B release context and is included for base-model reference. It is not a benchmark of this fine-tuned checkpoint unless explicitly stated in the fine-tune benchmark table above.

Category	Benchmark	Qwen3.5-27B	Gemma4-31B	Qwen3.5-35BA3B	Gemma4-26BA4B	Qwen3.6-35BA3B
Coding Agent	SWE-bench Verified	75.0	52.0	70.0	17.4	73.4
Coding Agent	SWE-bench Multilingual	69.3	51.7	60.3	17.3	67.2
Coding Agent	SWE-bench Pro	51.2	35.7	44.6	13.8	49.5
Coding Agent	Terminal-Bench 2.0	41.6	42.9	40.5	34.2	51.5
Coding Agent	Claw-Eval Avg	64.3	48.5	65.4	58.8	68.7
Coding Agent	Claw-Eval Pass^3	46.2	25.0	51.0	28.0	50.0
Coding Agent	SkillsBench Avg5	27.2	23.6	4.4	12.3	28.7
Coding Agent	QwenClawBench	52.2	41.7	47.7	38.7	52.6
Coding Agent	NL2Repo	27.3	15.5	20.5	11.6	29.4
Coding Agent	QwenWebBench	1068	1197	978	1178	1397
General Agent	TAU3-Bench	68.4	67.5	68.9	59.0	67.2
General Agent	VITA-Bench	41.8	43.0	29.1	36.9	35.6
General Agent	DeepPlanning	22.6	24.0	22.8	16.2	25.9
General Agent	Tool Decathlon	31.5	21.2	28.7	12.0	26.9
General Agent	MCPMark	36.3	18.1	27.0	14.2	37.0
General Agent	MCP-Atlas	68.4	57.2	62.4	50.0	62.8
General Agent	WideSearch	66.4	35.2	59.1	38.3	60.1
Knowledge	MMLU-Pro	86.1	85.2	85.3	82.6	85.2
Knowledge	MMLU-Redux	93.2	93.7	93.3	92.7	93.3
Knowledge	SuperGPQA	65.6	65.7	63.4	61.4	64.7
Knowledge	C-Eval	90.5	82.6	90.2	82.5	90.0
STEM & Reasoning	GPQA	85.5	84.3	84.2	82.3	86.0
STEM & Reasoning	HLE	24.3	19.5	22.4	8.7	21.4
STEM & Reasoning	LiveCodeBench v6	80.7	80.0	74.6	77.1	80.4
STEM & Reasoning	HMMT Feb 25	92.0	88.7	89.0	91.7	90.7
STEM & Reasoning	HMMT Nov 25	89.8	87.5	89.2	87.5	89.1
STEM & Reasoning	HMMT Feb 26	84.3	77.2	78.7	79.0	83.6
STEM & Reasoning	IMOAnswerBench	79.9	74.5	76.8	74.3	78.9
STEM & Reasoning	AIME26	92.6	89.2	91.0	88.3	92.7

Notes from the upstream Qwen3.6 release:

• SWE-Bench Series: internal agent scaffold with bash and file-edit tools; temp=1.0, top_p=0.95, 200K context window.
• Terminal-Bench 2.0: Harbor/Terminus-2 harness; 3h timeout, 32 CPU/48 GB RAM; temp=1.0, top_p=0.95, top_k=20, max_tokens=80K, 256K ctx; average of 5 runs.
• SkillsBench: evaluated via OpenCode on 78 tasks, using a self-contained subset excluding API-dependent tasks; average of 5 runs.
• NL2Repo: evaluated via Claude Code for other models, with temp=1.0, top_p=0.95, max_turns=900.
• QwenClawBench: internal real-user-distribution Claw agent benchmark; temp=0.6, 256K ctx.
• QwenWebBench: internal front-end code generation benchmark; bilingual EN/CN, seven categories, auto-render plus multimodal judge, BT/Elo rating system.
• TAU3-Bench: official user model with gpt-5.2 low reasoning effort and default BM25 retrieval.
• VITA-Bench: average subdomain scores, using claude-4-sonnet as judge.
• MCPMark: GitHub MCP v0.30.3, Playwright responses truncated at 32K tokens.
• MCP-Atlas: public set score, gemini-2.5-pro judge.
• AIME 26: full AIME 2026 I and II.

Training Pipeline

Qwen/Qwen3.6-35B-A3B
  -> supervised fine-tuning with LoRA
  -> merged full model
  -> Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled

Training configuration:

Setting	Value
Fine-tuning method	Supervised fine-tuning with LoRA
LoRA target	Attention-only modules
LoRA rank / alpha	32 / 32
Micro-batch size	1
Gradient accumulation	32
Epochs	2
Completed steps	762 / 762
Final reported training loss	0.3362497625740494
Dataset max tokens	8192
Max sequence length	32768

Training Data

The recipe samples and normalizes reasoning conversations from three datasets, then renders them with the qwen3-thinking chat template and response-only SFT masking.

Dataset	Requested sample count	Role
nohurry/Opus-4.6-Reasoning-3000x-filtered	3,900	Claude Opus reasoning trajectories
Jackrong/Qwen3.5-reasoning-700x	700	Curated Qwen reasoning samples
Roman1111111/claude-opus-4.6-10000x	9,633	Additional Claude Opus reasoning examples

Intended Use

This model is intended for reasoning-heavy text workflows such as coding assistance, planning, math-style reasoning, and structured analytical responses. Because the fine-tune is text-only, image/video behavior should be treated as inherited from the base model rather than improved by this training run.

Acknowledgements

Thanks to the Qwen team for the base model, Unsloth for the training stack, and Jackrong for the public reasoning-distillation workflow that inspired this fine-tune.