Qwen3-Coder-30B-A3B-Imatrix-GGUF

Qwen3-Coder-30B-A3B Imatrix GGUF

This repository provides a Q4_K_M GGUF quantization of Qwen3-Coder-30B-A3B-Instruct for daily coding and code-agent style workflows.

The quantization uses an importance matrix built from SWE-bench Lite examples. The calibration text includes issue reports, failing test hints, reference patch patterns, and test changes. The goal is to keep useful behavior for everyday development tasks such as debugging, patch planning, repository navigation, and test-focused edits.

This is a quantized conversion, not a fine-tuned model.

Usage

Example with llama.cpp:

llama-cli \
  -m qwen3-coder-30b-a3b-imatrix-q4_k_m.gguf \
  -ngl 999 \
  -c 4096 \
  -p "You are a coding assistant. Review this failing Python test and suggest a concise fix plan."

Details

• Base model: Qwen/Qwen3-Coder-30B-A3B-Instruct
• Format: GGUF
• Quantization: Q4_K_M
• Calibration method: llama.cpp importance matrix
• Calibration source: SWE-bench Lite derived prompts
• Target use: daily coding, repository debugging, patch planning, and code-agent prompts

Files

• qwen3-coder-30b-a3b-imatrix-q4_k_m.gguf: quantized model
• imatrix-daily-code.dat: importance matrix used for quantization
• daily_code_calibration.txt: calibration text used to build the importance matrix

Smoke Test

A short llama.cpp smoke test was run after quantization. The model loaded successfully and produced a coherent debugging plan for a Python parser test failure.

Example speed on A100 80GB PCIe:

• prompt: 1.2 tokens per second
• generation: 23.6 tokens per second

Reproducibility

The included importance matrix and calibration text are provided for reproducibility. They document the daily coding calibration used before quantization.

Notes

The importance matrix was generated for a focused daily coding workload. It is intended to be useful for common code repair and test analysis prompts rather than as a general benchmark claim.