MiniMax-M3-REAP32-Coder

<p align="center"><img src="./vmlx-logo.png" alt="vMLX" width="150"></p> <h1 align="center">MiniMax-M3-REAP32-Coder</h1> <p align="center"><b>A JANG-quantized MiniMax-M3 — coding/agentic + multimodal — for the <a href="https://mlx.studio">vMLX</a> engine (Apple Silicon / MLX).</b></p>

⚠️ Requires vMLX engine v1.5.67 or newer. This is a JANG-format model (JANG affine-mixed + AWQ quantization, REAP expert pruning, and the MiniMax-M3 MSA / Lightning-Indexer runtime). It will NOT load with transformers, vLLM, or generic MLX loaders — it needs vMLX's JANG loader + the M3 runtime. Coder support lands in vMLX ≥ 1.5.67.

What is a JANG model?

JANG is vMLX's quantization + packing format: mixed-precision affine quantization (per-projection bit widths) + AWQ activation-aware scaling + REAP expert pruning, described by a jang_config.json. Weights stay quantized in GPU memory and are loaded by vMLX's JANG loader. Because the format and the MiniMax-M3 runtime (MSA dual-cache, Lightning Indexer, partial RoPE, vision tower) are vMLX-specific, these models run only on vMLX ≥ 1.5.67.

Run it

1. Install/update vMLX 1.5.67+ — https://mlx.studio (or pip install -U vmlx).
2. App: Server → New Session → pick/download this model → Start → chat.
3. CLI: vmlx-engine serve JANGQ-AI/MiniMax-M3-REAP32-Coder --reasoning-parser minimax_m3 --tool-call-parser minimax_m3

Highlights

• Coding: HumanEval pass@1 = 100% (82/82 on a scrambled half of HumanEval, first-sample) — pass@5 = 1.000.
• Arithmetic/reasoning recovered vs the base REAP quant (with reasoning enabled): ~6/7 on a 7-task probe.
• Multimodal (vision) kept. ~105 GB on disk.

Build

• Base: MiniMaxAI/MiniMax-M3 (60 layers, MoE, MSA Lightning Indexer, GQA, partial RoPE).
• REAP pruning: keep 87/128 routed experts per MoE layer (32% pruned), saliency-scored.
• JANG affine quant (group_size 64): routed gate/up = 2-bit + AWQ pre-scaling, down = 3-bit; shared experts 6-bit; attention 8-bit; embeddings 6-bit; lm_head 8-bit; Lightning Indexer + norms FP16; vision 8-bit.
• "Floor" expert recipe: protect the proven coding experts (coding saliency) + add top math experts, so coding stays intact while math improves.
• Calibration: Vera (agentic-coder) dominant + GSM8K (math reasoning).

Attribution

• Base model: MiniMaxAI/MiniMax-M3
• Expert pruning: REAP (Cerebras, ICLR 2026, arXiv:2510.13999)
• Vera agentic-coder calibration dataset + evaluation/testing: @hornsman1 (hornsan1 on GitHub)
• Additional math-reasoning calibration: GSM8K
• Quantization & runtime: JANG / vMLX

Credits

• Vera dataset & model testing: @hornsman1 (hornsan1 on GitHub)