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

TEXT GENERATIONConcurrent Unit Cost:3Model Size:35.1BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Apr 17, 2026License:apache-2.0Architecture:Transformer0.1K Open Weights Featherless Exclusive Cold

hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled is a 35.1 billion parameter instruction-tuned causal language model developed by hesamation. It is a fine-tune of Qwen/Qwen3.6-35B-A3B, specifically optimized for structured Claude Opus-style reasoning traces and stable long-form problem solving through chain-of-thought (CoT) distillation. This model excels in reasoning-heavy text workflows such as coding assistance, planning, and mathematical reasoning, leveraging its 32768 token context length.

Loading preview...

Model Overview

hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled is a 35.1 billion parameter model fine-tuned from the Qwen/Qwen3.6-35B-A3B base. Its primary goal is to enhance reasoning capabilities by distilling chain-of-thought (CoT) traces, predominantly sourced from Claude Opus 4.6. This process aims to maintain the base model's strong agentic coding and reasoning foundation while integrating a more structured, Claude Opus-style approach to problem-solving.

Key Capabilities

  • Enhanced Reasoning: Specialized in generating structured, Claude Opus-style reasoning traces for complex problems.
  • Agentic Coding: Inherits and refines the base Qwen3.6's proficiency in agentic coding, including frontend workflows and repository-level reasoning.
  • Long-form Problem Solving: Designed for more stable and effective long-form problem resolution.
  • MMLU-Pro Performance: Achieved a 75.71% score on MMLU-Pro overall in a comparative smoke test, a +32.85 percentage point improvement over its base model's 42.86% on the same limited benchmark.

Training Details

The model underwent supervised fine-tuning with LoRA, focusing on attention-only modules. The training data primarily consisted of reasoning conversations from three datasets: nohurry/Opus-4.6-Reasoning-3000x-filtered, Jackrong/Qwen3.5-reasoning-700x, and Roman1111111/claude-opus-4.6-10000x, all rendered with the qwen3-thinking chat template and response-only SFT masking. The fine-tuning was text-only, meaning its vision capabilities are inherited from the base model and not further trained.

Intended Use Cases

This model is best suited for:

  • Coding assistance: Particularly for agentic coding tasks.
  • Planning and Strategy: Generating structured plans and analytical responses.
  • Mathematical Reasoning: Solving complex math-style problems.
  • Structured Analytical Responses: Producing detailed, step-by-step reasoning.

Finetunes
2 models