clzoro/Qwen3.6-27B-Claude-Distill-v2

VISIONConcurrency Cost:2Model Size:27BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Apr 26, 2026License:apache-2.0Architecture:Transformer Open Weights Cold

clzoro/Qwen3.6-27B-Claude-Distill-v2 is a 27 billion parameter language model, a supervised fine-tune of Qwen3.6-27B by clzoro. It is specifically enhanced for instruction-following and reasoning, trained on a large dataset of Claude-generated conversations. This model excels particularly in mathematical and coding tasks, which constitute approximately 80% of its training data, while maintaining the base model's 32K token context length.

Loading preview...

Model Overview

clzoro/Qwen3.6-27B-Claude-Distill-v2 is a 27 billion parameter language model, built upon the Qwen3.6-27B base model. It has undergone full supervised fine-tuning (SFT) using a substantial dataset of 125,175 Claude-distilled conversation pairs. This training process aims to significantly enhance the model's instruction-following and reasoning capabilities, inheriting the strengths of Claude-generated data.

Key Capabilities & Differentiators

  • Enhanced Instruction Following: Improved ability to understand and execute complex instructions due to SFT on high-quality Claude-distilled data.
  • Strong Math and Code Performance: The training dataset is heavily weighted towards mathematical (65.5%) and coding (15.1%) tasks, making the model particularly proficient in these domains.
  • Reasoning Focus: Significant portion of training data dedicated to reasoning, contributing to its ability to handle complex problem-solving.
  • "Thinking Mode" Feature: By default, the model operates in a "thinking mode" that generates internal reasoning steps before producing a final response, which can be disabled for direct answers.
  • Qwen3.6 Base: Benefits from the robust capabilities and performance of the underlying Qwen3.6-27B architecture.

Ideal Use Cases

  • Complex Mathematical Problem Solving: Excels in tasks requiring detailed mathematical reasoning and solutions.
  • Code Generation and Analysis: Highly effective for generating code, debugging, and understanding programming logic.
  • Instruction-Following Applications: Suitable for scenarios where precise adherence to instructions is critical.
  • Reasoning-Intensive Tasks: Can be applied to problems demanding logical deduction and step-by-step reasoning.

Limitations

  • Primarily trained on English and Chinese data, potentially limiting performance in other languages.
  • The strong bias towards math and code in training data may affect performance in less represented domains.
  • As a distilled model, it may carry biases from the Claude-generated training data.
  • The model has not undergone RLHF or similar safety alignment.

Finetunes
1 models