khazarai/Qwen3-4B-Kimi2.5-Reasoning-Distilled

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:4BQuant:BF16Ctx Length:32kPublished:Mar 21, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Warm

khazarai/Qwen3-4B-Kimi2.5-Reasoning-Distilled is a 4 billion parameter language model developed by khazarai, fine-tuned from the Qwen3-4b-Thinking-2507 base model. Optimized for structured, long-form reasoning, it bridges the gap between small, efficient models and complex reasoning capabilities. This model excels at problem decomposition, self-correction, and providing detailed analytical answers, making it suitable for tasks requiring advanced logical processing.

Loading preview...

Model Overview

khazarai/Qwen3-4B-Kimi2.5-Reasoning-Distilled is a 4 billion parameter language model developed by khazarai, specifically optimized for structured, long-form reasoning tasks. It is fine-tuned from the Qwen3-4b-Thinking-2507 base model using a specialized distillation dataset generated by Kimi-2.5-thinking.

Key Capabilities

  • Enhanced Reasoning: Designed to bridge the gap between smaller, efficient models and the complex reasoning typically found in much larger models.
  • Problem Decomposition: Excels at breaking down complex problems into manageable parts.
  • Self-Correction: Capable of self-correcting its reasoning process to improve accuracy.
  • Detailed Analysis: Provides comprehensive and analytical answers.
  • Performance: Achieves a score of 76.09 on the khazarai/Multi-Domain-Reasoning-Benchmark, outperforming its base model Qwen/Qwen3-4B-Thinking-2507 (73.73).

Training Details

The model was fine-tuned using Unsloth + QLoRa on the khazarai/kimi-2.5-high-reasoning-250x dataset. This dataset comprises 250 samples and 1,114,407 tokens, with Kimi-2.5-Thinking serving as the teacher model.

Good For

  • Applications requiring advanced logical reasoning and problem-solving.
  • Generating detailed analytical responses.
  • Use cases where efficient, smaller models need to perform complex reasoning tasks.