Qwen2.5-0.5B-ReasonChat

Qwen2.5-0.5B-ReasonChat

A compact 373MB model that combines reasoning capability with natural conversational ability. Built by merging two independently fine-tuned variants of Qwen2.5-0.5B-Instruct using task arithmetic, then quantized to Q4_K_M for edge deployment.

Model Details

• Base Model: Qwen/Qwen2.5-0.5B-Instruct
• Parameters: 502M total
• Quantization: Q4_K_M (GGUF)
• File Size: ~373MB
• RAM Required: ~500MB
• Format: GGUF (llama.cpp / Ollama compatible)
• Language: English
• License: Apache 2.0

Training Overview

This model was built through a multi-stage process:

Stage 1 — Reasoning Distillation Fine-tuned on 116k reasoning samples from Claude 4.6 Opus traces and DeepSeek-R1 generated chain-of-thought data. Config: LoRA r=128, lr=2e-4, 3 epochs, seq=4096.

Stage 2 — General + Factual + Anti-loop Fresh fine-tune on 70k mixed samples covering general chat, factual QA, math, coding, and anti-loop correction pairs. Config: LoRA r=16, lr=7e-5, 1 epoch, seq=1024.

Stage 3 — Merge Both models merged using task arithmetic (60% reasoning + 40% general) via mergekit to combine strengths of both stages.

Full Dataset Mix

Total: ~199k samples across both training stages

Hardware

Trained on AMD Instinct MI300X (192GB VRAM), ROCm 6.2.4, Unsloth 2026.3.3, PyTorch 2.7.1+rocm6.2.4.

Usage

Ollama

ollama run hf.co/lqfdjbf32n/Qwen2.5-0.5B-ReasonChat:Q4_K_M

llama.cpp

llama-cli -m qwen0.5b-merged-v2v3-Q4_K_M.gguf \
    -p "<|im_start|>system\nYou are a helpful, concise assistant. Answer directly and stop when done.<|im_end|>\n<|im_start|>user\nYour question here<|im_end|>\n<|im_start|>assistant\n" \
    -n 256

Strengths

• Natural casual conversation without overthinking
• Basic math and reasoning capability
• Factual QA accuracy
• Anti-loop trained — stops when done
• Fits in ~500MB RAM — suitable for edge devices

Limitations

• Small model (0.5B parameters) — complex multi-step reasoning may fail
• Primarily trained on English data
• Not suitable for production use without validation