PKSGIN/qwen3-30b-selective-quant

Qwen3-30B-A3B Selective Quantization for Security & Code Analysis

A selectively quantized version of Qwen3-30B-A3B optimized for local inference on Apple Silicon, with preserved precision on code analysis experts.

MLX Model Here :

PKSGIN/qwen3-30b-selective-quant-mlx

Model Description

This model applies heterogeneous quantization to Qwen3-30B-A3B's Mixture-of-Experts architecture:

• 9 coding experts → Q8 (8-bit precision)
• 119 general experts → Q4 (4-bit precision)
• Router, attention, embeddings → FP16 (full precision)

The result: 18 GB on disk, runs at 52 tokens/sec on MacBook Pro M4 (32GB), with materially better code analysis quality than uniform Q4.

Why Selective Quantization?

Standard quantization tools (mlx_lm.convert, llama.cpp) compress all experts uniformly. For security and coding workloads, this degrades the exact experts you need most.

This model was profiled on 200 coding/security prompts to identify which experts activate for:

• Infrastructure as Code (Terraform, Kubernetes)
• Code review and vulnerability analysis
• Security alert triage

Those 9 experts are preserved at Q8 while compressing the rest to Q4.

Target Use Cases

Local inference for sensitive workloads:

• Infrastructure as Code review without cloud API calls
• Code security analysis in air-gapped environments
• SIEM alert triage with internal IP/hostname context
• Pull request review for injection vulnerabilities

Zero data exfiltration. Zero tokens leaving your device.

Hardware Requirements

• MacBook Pro M4 with 32GB unified memory (tested)
• MacBook Pro M3/M2 with 32GB+ will work
• Peak memory: ~18 GB during inference
• Inference speed: 52 tokens/sec (M4 Pro)

Usage

# Install MLX
pip install mlx mlx-lm

# Run inference
python -m mlx_lm.generate \
    --model PKSGIN/qwen3-30b-selective-quant \
    --prompt "Review this Terraform for security issues: ..."

Performance

Metric	Value
Model size on disk	18 GB
Peak memory	18.1 GB
Tokens/sec (M4 Pro)	52
Prompt processing	4.4 tokens/sec
Compression ratio	3.0x vs FP16

Quantization Details

Profiled experts (coding ratio > 0.65):

[21, 27, 31, 43, 59, 66, 71, 113, 126]

Memory breakdown:

• FP16 components (router/embeddings): 1.2 GB
• Q8 coding experts: 2.1 GB
• Q4 other experts: 16.7 GB
• Total: 20 GB weights → 18 GB after MLX conversion

Limitations

• Apple Silicon only — requires MLX framework
• English-centric profiling — coding experts identified from English prompts
• Domain-specific — optimized for code/security, may underperform on creative writing
• Not benchmarked — no formal comparison vs uniform Q4 on CyberSecEval/HumanEval yet

Training & Profiling

1. Base model: Qwen/Qwen3-30B-A3B
2. Profiling: 200 coding/security prompts + 200 general prompts on A100 80GB
3. Quantization: Selective Q8/Q4 applied via custom pipeline
4. Conversion: Dequantized → MLX switch_mlp format → final Q4

Full pipeline: 4 scripts, ~$2 GPU rental, 72 minutes total.

Citation

@misc{qwen3-selective-quant-2026,
  author = {Prasanna Kanagasabai},
  title = {Qwen3-30B-A3B Selective Quantization for Security Workloads},
  year = {2026},
  publisher = {HuggingFace},
  url = {https://huggingface.co/PKSGIN/qwen3-30b-selective-quant}
}

License

Same as base model: Qwen License

Contact

Built by PK (@PKSGIN) — CISO, Singapore. Email : prasanna.in@gmail.com