qwen3-4b-reasoning is a 4B-parameter Qwen3-based reasoning “backfill” fine-tune (joeyzero/Qwen3-4B-Reasoning-Backfill-v0.1) converted to GGUF for llama.cpp/Ollama, with ~40K context and published as Q4_K_M (recommended) and iq4_xs (smaller).

Qwen3-4B-Reasoning

A compact 4B-parameter Qwen3 reasoning model that reconstructs explicit chain-of-thought, packaged as GGUF for fast local inference with Ollama and llama.cpp.

🚀 Overview

This model is a GGUF conversion of joeyzero/Qwen3-4B-Reasoning-Backfill-v0.1, an experimental reasoning fine-tune built on Qwen/Qwen3-4B. It was trained to “backfill” reasoning, reconstructing a plausible chain of thought that connects a user instruction to a known solution while preserving the original answer. The result is a small, efficient model that thinks step by step before responding, making it well suited for adding explicit reasoning traces and for lightweight local reasoning tasks.

🎯 Key Features

Step-by-Step Reasoning: Generates explicit thinking traces before producing a final answer
Reasoning Backfill: Trained to reconstruct plausible chains of reasoning for instruction/solution pairs
Compact & Efficient: At 4B parameters with 4-bit quantization (~2.3–2.5 GB), it runs comfortably on modest hardware
Full Qwen3 Foundation: Built on Qwen3-4B, retaining its multilingual and general capabilities
40K Context: Configured with a 40,960-token context window for longer prompts and traces

🏷️ Available Versions

Tag	Size	BPW	Notes
`latest` / `Q4_K_M`	2.5 GB	~4.85	Recommended balance of quality and size
`iq4_xs`	2.3 GB	~4.25	Smaller, for low VRAM

# Examples
ollama run richardyoung/qwen3-4b-reasoning              # Recommended (Q4_K_M)
ollama run richardyoung/qwen3-4b-reasoning:iq4_xs       # Smaller

💻 Quick Start

# Pull and run
ollama run richardyoung/qwen3-4b-reasoning

# Or with a specific version
ollama run richardyoung/qwen3-4b-reasoning:Q4_K_M

Recommended sampling parameters: temperature=0.6, top_p=0.95, top_k=20.

🛠️ Use Cases

Reasoning Augmentation: Backfill explicit chain-of-thought traces for datasets that lack them
Step-by-Step Problem Solving: Math, logic, and structured reasoning tasks
Local Reasoning: Run a thinking model entirely on-device with low resource usage
Education: Inspect how a model reasons toward a known answer

📋 System Requirements

VRAM	Performance
4 GB	Good performance
6 GB+	Excellent performance
CPU-only	Usable; slower generation

🔧 Technical Details

Base Model: joeyzero/Qwen3-4B-Reasoning-Backfill-v0.1 (built on Qwen/Qwen3-4B) Parameters: ~4B Context Length: 40,960 tokens Prompt Format: ChatML Architecture: qwen3 Quantization: Q4_K_M (~4.85 bits per weight) / IQ4_XS (~4.25 bits per weight) License: Apache 2.0

⚠️ Disclaimer

This is an experimental reasoning model. Generated chains of thought are reconstructions and may not reflect a unique or strictly correct reasoning path; always verify outputs for important tasks. Use responsibly and in accordance with applicable laws and regulations.

🙏 Acknowledgments

Base Model: joeyzero/Qwen3-4B-Reasoning-Backfill-v0.1
Foundation: Qwen/Qwen3-4B by Alibaba
Quantization: llama.cpp

Built & maintained by Richard Young · DeepNeuro