empero-ai/Qwen3.5-9B-Claude-Code-GGUF

CodeClawd — Qwen3.5-9B Claude Code (GGUF)

GGUF quantizations of empero-ai/Qwen3.5-9B-Claude-Code — A supervised fine-tune of empero-ai/Qwen3.5-9B-Claude-Opus-4.6-Distill on real Claude Code / Codex agentic sessions from the DataClaw dataset collection.

The model is trained to behave like a capable coding agent: thinking through problems, calling tools, interpreting results, and responding clearly — mirroring the full agentic loop of a real Claude Code session.

The model thinks with <think> tags, calls tools with XML-style <tool_call> blocks.

---

Available Quantizations

Filename	Quant	Size	RAM Required	Quality
Qwen3.5-9B-Claude-Code-Q2_K.gguf	Q2_K	~3.6 GB	~6 GB	Lowest — not recommended
Qwen3.5-9B-Claude-Code-Q3_K_M.gguf	Q3_K_M	~4.5 GB	~7 GB	Low
Qwen3.5-9B-Claude-Code-Q4_0.gguf	Q4_0	~5.2 GB	~8 GB	Acceptable
Qwen3.5-9B-Claude-Code-Q4_K_M.gguf	Q4_K_M	~5.4 GB	~8 GB	Recommended — best size/quality tradeoff
Qwen3.5-9B-Claude-Code-Q5_K_M.gguf	Q5_K_M	~6.4 GB	~9 GB	High
Qwen3.5-9B-Claude-Code-Q6_K.gguf	Q6_K	~7.3 GB	~10 GB	Very high
Qwen3.5-9B-Claude-Code-Q8_0.gguf	Q8_0	~9.5 GB	~12 GB	Near-lossless
Qwen3.5-9B-Claude-Code-F16.gguf	F16	~18 GB	~22 GB	Lossless

Recommended: Q4_K_M for most users. Q5_K_M or Q6_K if you have extra VRAM and want higher quality reasoning.

---

Model Details

Property	Value
Original model	empero-ai/Qwen3.5-9B-Claude-Code
Base model	empero-ai/Qwen3.5-9B-Claude-Opus-4.6-Distill
Fine-tune type	SFT (LoRA r=64/alpha=128, merged)
Parameters	9B
Training data	DataClaw — 29 datasets of real Claude Code / Codex sessions
Context length	4096 tokens
License	Apache 2.0

---

Usage with llama.cpp

# Basic inference
./llama-cli \
  -m Qwen3.5-9B-Claude-Code-Q4_K_M.gguf \
  -n 1024 \
  --temp 0.7 \
  --top-p 0.9 \
  -p "<|im_start|>system
You are an expert AI coding assistant. You help users with software engineering tasks including writing code, debugging, refactoring, explaining code, and more. You think through problems carefully inside <think> tags before acting. You have access to tools like Read, Edit, Write, Bash, Grep, Glob, and others to interact with the user's codebase. You use tools when needed and provide clear, concise responses.<|im_end|>
<|im_start|>user
Read the file main.py and explain what it does.<|im_end|>
<|im_start|>assistant
"

# As an OpenAI-compatible server
./llama-server \
  -m Qwen3.5-9B-Claude-Code-Q4_K_M.gguf \
  --host 0.0.0.0 \
  --port 8080 \
  -c 4096 \
  -ngl 99

---

Usage with Ollama

Create a Modelfile:

FROM Qwen3.5-9B-Claude-Code-Q4_K_M.gguf

SYSTEM """You are an expert AI coding assistant. You help users with software engineering tasks including writing code, debugging, refactoring, explaining code, and more. You think through problems carefully inside <think> tags before acting. You have access to tools like Read, Edit, Write, Bash, Grep, Glob, and others to interact with the user's codebase. You use tools when needed and provide clear, concise responses. You show your reasoning process and use tools methodically to accomplish tasks."""

PARAMETER temperature 0.7
PARAMETER top_p 0.9
PARAMETER num_ctx 4096

ollama create codeclawd -f Modelfile
ollama run codeclawd

---

Usage with LM Studio

1. Download the .gguf file of your chosen quantization
2. Open LM Studio → Load Model → select the file
3. In the system prompt field, paste:

You are an expert AI coding assistant. You help users with software engineering tasks including writing code, debugging, refactoring, explaining code, and more. You think through problems carefully inside <think> tags before acting. You have access to tools like Read, Edit, Write, Bash, Grep, Glob, and others to interact with the user's codebase. You use tools when needed and provide clear, concise responses. You show your reasoning process and use tools methodically to accomplish tasks.

4. Set temperature: 0.7, top-p: 0.9, context: 4096

---

Agentic Format

This model is trained on the full agentic loop:

user → <think>...</think> → <tool_call> → <tool_result> → response → repeat

Tool Call Format

<tool_call>
<tool>Read</tool>
<input>
{"file_path": "/path/to/file.py"}
</input>
</tool_call>

<tool_result>
<tool>Read</tool>
<output>
... file contents ...
</output>
</tool_result>

Thinking Format

<think>
The user wants to refactor the auth module. I should first read the existing
implementation before suggesting changes.
</think>

---

Intended Use

• Local agentic coding assistants on consumer hardware
• IDE integrations (Continue, Cursor, etc.) running locally
• Offline / air-gapped development environments
• Research into agentic SFT distillation from real sessions

Limitations

• Tool format is XML-based — not OpenAI function-call JSON
• SFT only (no RLHF or DPO)
• May hallucinate tool outputs if not connected to a real tool executor
• Quantization below Q4 may degrade reasoning quality noticeably

---

Original Model

• HF (safetensors): empero-ai/Qwen3.5-9B-Claude-Code

License

Apache 2.0 — see base model for details.