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

GGUF quantizations of hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled, a reasoning SFT fine-tune of Qwen/Qwen3.6-35B-A3B on Claude Opus 4.6-style chain-of-thought distillation data.

The source fine-tune 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. Treat these GGUF files as text-generation/runtime quantizations of the merged fine-tuned checkpoint.

• Developed by: @hesamation
• Source model: hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled
• 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.

Available GGUF Quantizations

This repo is intended to host the following GGUF variants. Files are uploaded as each quantization finishes.

Quant	Typical use
Q4_K_M	Smallest practical general-purpose quant for local inference
Q5_K_M	Better quality/size balance than Q4
Q6_K	Higher-quality quant when VRAM/RAM budget allows
Q8_0	Largest quant here; closest to source quality among these options

Benchmark Results

The benchmark below was run on the merged source model, not separately on each GGUF quant. Quantization can change scores, especially at lower bitrates, so treat this as source-checkpoint context.

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	Source 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. Source merged model: hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled.

Community benchmarks welcome

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

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

Training Summary

Qwen/Qwen3.6-35B-A3B
  -> supervised fine-tuning with LoRA
  -> merged full model
  -> GGUF quantization with llama.cpp

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 source model 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

These GGUF files are intended for local or server-side text inference through runtimes that support GGUF and the Qwen3.6 architecture, such as recent llama.cpp builds. Choose the quantization based on your memory budget and quality target.

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, llama.cpp for GGUF tooling, and Jackrong for the public reasoning-distillation workflow that inspired this fine-tune.