---
tags:
- gguf
- llama.cpp
- unsloth
- reasoning
- distillation
- sft
license: apache-2.0
datasets:
- khazarai/kimi-2.5-high-reasoning-250x
language:
- en
base_model:
- khazarai/Qwen3-4B-Kimi2.5-Reasoning-Distilled
pipeline_tag: text-generation
---

# Qwen3-4B-Kimi2.5-Reasoning-Distilled : GGUF

## Model: khazarai/Qwen3-4B-Kimi2.5-Reasoning-Distilled

![alt="General Benchmark Comparison Chart"](benchmark/Kimi_distilled.png)

- **Success Rate**: 76.09%

## Model: Qwen/Qwen3-4B-Thinking-2507

![alt="General Benchmark Comparison Chart"](benchmark/BaseModel.png)

- **Success Rate**: 73.73%

- **Benchmark**: khazarai/Multi-Domain-Reasoning-Benchmark
- **Total Questions**: 100

`Qwen3-4B-Kimi2.5-Reasoning-Distilled` is a fine-tuned language model optimized for structured, long-form reasoning. It is derived from the Qwen3-4b-Thinking-2507 base model and fine-tuned using a specialized distillation dataset generated by Kimi-2.5-thinking.

This model is designed to bridge the gap between small, efficient models (0.6B–4B range) and the complex reasoning capabilities typically found in much larger models. It excels at breaking down problems, self-correcting, and providing detailed analytical answers.

**Base Model**:	Qwen3-4b-Thinking-2507

**Training Technique**:	Unsloth + QLoRa
 

## Available Model files:
- `qwen3-4b-thinking-2507.BF16.gguf`
- `qwen3-4b-thinking-2507.Q8_0.gguf`
- `qwen3-4b-thinking-2507.Q6_K.gguf`
- `qwen3-4b-thinking-2507.Q4_K_M.gguf`

## Ollama
An Ollama Modelfile is included for easy deployment.
 

## Provided Quants

(sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants)

| Type | Size/GB | Notes |
|:-----|--------:|:------|
| Q4_K_M | 2.5 | fast, recommended |
| Q6_K | 3.3 | very good quality |
| Q8_0 | 4.2 | fast, best quality |
| f16 | 8.0 | 16 bpw, overkill |

Here is a handy graph by ikawrakow comparing some lower-quality quant
types (lower is better):

![image.png](https://www.nethype.de/huggingface_embed/quantpplgraph.png)

And here are Artefact2's thoughts on the matter:
https://gist.github.com/Artefact2/b5f810600771265fc1e39442288e8ec9
 

## Dataset

The model was fine-tuned on the [khazarai/kimi-2.5-high-reasoning-250x](https://huggingface.co/datasets/khazarai/kimi-2.5-high-reasoning-250x)  

Dataset Composition:
 - Total Samples: 250
 - Total Tokens: 1,114,407
 - Teacher Model: Kimi-2.5-Thinking

## Acknowledgements

**Unsloth** for the incredibly fast and memory-efficient training framework.