rogermaiocchi/SimpleSD-30B-instruct-MLX-4bit

SimpleSD-30B-instruct — Derivative MLX 4-bit

Apple Machine Learning Research Model is licensed under the Apple Machine Learning Research Model License Agreement.

What this repository is

This is a Model Derivative of apple/SimpleSD-30B-instruct.

• Upstream model: apple/SimpleSD-30B-instruct
• Upstream revision: 762e318a68c2792125dabc7ce220da6d630e0086
• Base architecture: Qwen3MoeForCausalLM (Qwen3-30B-A3B-Instruct-2507 lineage)
• Derivative operation: quantization only (MLX affine quantization, 4 bits, group size 64, router/gate at 8 bits; effective 4.502 bits per weight)
• Training performed: none. No weight was retrained, fine-tuned or otherwise trained.
• Conversion tool: mlx_lm.convert (mlx-lm 0.31.3 / mlx 0.32.2)

Disclosure of modifications and changes

As required by the Apple Machine Learning Research Model License Agreement, the modifications made to the upstream Apple Machine Learning Research Model are disclosed here in full:

1. The bfloat16 weights were quantized to 4-bit affine MLX format with group size 64 (gate/router projections kept at 8-bit). No other numerical transformation was applied.
2. The upstream LICENSE file is included unchanged.
3. The upstream config.json was extended with the MLX quantization block that describes the conversion; the model architecture fields are unchanged from upstream.
4. The tokenizer and chat template are unchanged from upstream.

License and permitted use

This derivative is licensed under the Apple Machine Learning Research Model License, reproduced verbatim in the LICENSE file of this repository. That license restricts use, modification, redistribution and derivative creation to Research Purposes — meaning non-commercial scientific research and academic development. It does not cover any commercial exploitation, product development, or use in a commercial product or service.

The upstream LICENSE file must accompany any further redistribution of these weights or of derivatives of them, and any further derivative must again be clearly identified as such with disclosure of its own modifications.

Usage

pip install mlx-lm
mlx_lm.generate \
  --model rogermaiocchi/SimpleSD-30B-instruct-MLX-4bit \
  --prompt "Write a Python function that merges two sorted lists." \
  --max-tokens 512

Apple Silicon is required for the MLX runtime.

Provenance

This derivative was produced locally on an Apple Silicon host from the pinned upstream revision. The local conversion, its exit code, and the weight hashes are recorded in the operator's evidence store. See DERIVATIVE_NOTICE.md.