Qwen3-8B-MedReasonPath

Qwen3-8B fine-tuned for stronger step-by-step medical reasoning, built to retain its native thinking mode and its existing tool-calling ability.

This is a full standalone model. No PEFT, no Unsloth, no extra libraries required — load it exactly like any other transformers causal LM.

Base model

Qwen/Qwen3-8B — 8.2B parameters, hybrid thinking / non-thinking model with native agentic tool-calling support.

Training data

• MedReason (5,000 examples) — knowledge-graph grounded medical reasoning, for clinical reasoning quality.
• xLAM Function-Calling (681 examples) — general tool-calling data, mixed in during training to help preserve agentic behavior.

Method

QLoRA fine-tuning (4-bit, rank 16, alpha 32) on a single T4 GPU, using Unsloth purely as a training-time tool. The resulting adapter was merged into the base model and released here as a full fp16 checkpoint — Unsloth is not needed to use this model, only transformers.

Result

Improved step-by-step reasoning inside Qwen3-8B's native <think>...</think> mode on medical/clinical questions, while keeping thinking mode, non-thinking mode, and tool-calling all functional.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "Rumiii/Qwen3-8B-MedReasonPath", torch_dtype="auto", device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("Rumiii/Qwen3-8B-MedReasonPath")

messages = [{"role": "user", "content": "What causes aortic dissection?"}]
text = tokenizer.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True, enable_thinking=True
)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=600, temperature=0.6, top_p=0.95, top_k=20)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Set enable_thinking=False for direct, non-reasoning answers.

License

Apache 2.0, matching the base model.

Disclaimer

Research and educational project. Not validated for clinical use — not a substitute for professional medical advice.

GGUF

An 8-bit quantized GGUF version (Qwen3-8B-MedReasonPath-Q8_0.gguf, ~8.7GB) is also available in this repo, for use with llama.cpp, Ollama, or LM Studio.

# llama.cpp
./llama-cli -hf Rumiii/Qwen3-8B-MedReasonPath:Q8_0