ImTamsi/qwen3.6-27b-code-reviewer

Qwen3.6-27B Code Review — deployment configuration (no weights)

This repository does not contain model weights.
It documents how to serve Qwen/Qwen3.6-27B for the AI Code Reviewer MCP project: vLLM flags, generation settings, prompts, and the JSON output contract. Download weights from the base model repo.

Type	Deployment config & model card
Weights	Qwen/Qwen3.6-27B
Companion dataset	ImTamsi/code-review-findings-samples
Gradio demo	ImTamsi/ai-code-reviewer
Hackathon org	build-small-hackathon

What this repo is (and is not)

Is:

• A reproducible serving recipe (vLLM CLI, hardware table, generation params)
• The expected JSON schema for automated findings
• A Hub entry point for the MCP + Space ecosystem

Is not:

• A fine-tuned checkpoint — use the base Qwen model for weights
• A drop-in transformers weights folder — only README.md is published here

What the served model does

Given a GitHub repository's source (file tree + file contents, and optionally a pull request diff), the model produces structured findings across six analyses:

• Code review
• Potential bug detection
• Technical debt
• Security
• Performance
• Missing test suggestions

The full MCP server uses tool-calling (read_file, list_files). The Gradio Space uses a simplified single-pass flow with the same prompts.

Serving (vLLM, OpenAI-compatible)

vllm serve Qwen/Qwen3.6-27B \
  --port 8000 \
  --language-model-only \
  --max-model-len 32768 \
  --kv-cache-dtype fp8 \
  --reasoning-parser qwen3 \
  --enable-auto-tool-choice --tool-call-parser qwen3_coder \
  --api-key "$QWEN_API_KEY"

• --language-model-only disables the vision encoder (code review is text-only), freeing VRAM for KV cache.
• --tool-call-parser qwen3_coder + --enable-auto-tool-choice enable OpenAI-style function calling used by the agentic read_file / list_files tools.

Hardware

Quantization	VRAM	Example AWS instance
AWQ / INT4	~17 GB	g5.2xlarge (A10G 24 GB)
FP8	~28 GB	g6e.xlarge (L40S 48 GB)
BF16	~56 GB	g5.12xlarge (4x A10G, TP=4)

Generation settings

Parameter	Value
temperature	0.2
max_tokens	4096
response_format	json_object

Low temperature keeps findings deterministic and grounded.

Output contract

The model must return a single JSON object:

{
  "summary": "string",
  "findings": [
    {
      "severity": "critical | high | medium | low | info",
      "category": "string",
      "file": "string | null",
      "line": "number | null",
      "title": "string",
      "explanation": "string",
      "suggestion": "string"
    }
  ]
}

See ImTamsi/code-review-findings-samples for annotated examples.

Prompts

The system prompt and the six per-analysis prompts are published with the project under packages/mcp-server/src/prompts/.

Intended use and limitations

• Intended as an assistant for code review; findings are suggestions, not guarantees.
• The model can miss issues or report false positives. Always have a human confirm security- and correctness-critical findings.
• Large repositories are partially inlined within a context budget; the agentic read_file tool mitigates but does not eliminate coverage gaps.

License

Apache-2.0, matching the base model Qwen/Qwen3.6-27B.