JANGQ-AI/MiniMax-M3-REAP40-JANG_2L

MiniMax-M3 · REAP-40 · JANG_2L

⚠️ Requires vMLX ≥ v1.5.62

Earlier vMLX builds contain a runtime cache bug that causes repetition loops on long output. This is an engine issue, not a weights issue — update vMLX to v1.5.62 or later before running this model. On v1.5.62+ generation is clean.

A space-efficient MiniMax-M3 bundle for Apple Silicon: 40 % REAP expert pruning + JANG_2L mixed-precision quantization, ~95 GB — the smaller, lower-RAM sibling of REAP-32 (more pruning, less resident memory). Runs comfortably on a single 128 GB Mac via vMLX / MLX.

What this is

• Base: MiniMax-M3 (model_type=minimax_m3_vl) — MoE, GQA-4, MSA Lightning Indexer, vision tower.
• Pruning: REAP saliency pruning, 40 % of routed experts removed (77 of 128 kept per MoE layer), highest-saliency experts retained.
• Quantization (JANG_2L, affine, group size 64):

  tensor	bits
  routed experts gate_proj / up_proj	2
  routed experts down_proj	3
  shared experts	6
  dense MLP (layers 0–2)	6
  attention q/k/v/o	8
  embeddings	6
  lm_head	8
  vision tower + projectors	8
  norms, router gate, MSA indexer	fp16

  down_proj is kept at 3-bit (the rest of the routed experts are 2-bit) for stable long-form coherency. The full per-module bit map is written into config.json (quantization) and applied automatically by the loader.

Which one to use

• REAP-40 (this repo, ~95 GB): smaller, lower resident RAM, more headroom.
• REAP-32 (~105 GB): keeps more experts; higher quality ceiling, closer to the RAM limit on 128 GB machines.

Usage

Load in vMLX (v1.5.62+); the engine autodetects minimax_m3_vl and applies the correct settings (native MSA cache, paged cache off, per-module quant map). Sampling defaults ship in generation_config.json (temperature=1.0, top_p=0.95).

Attribution

• Quantization & packaging: Jinho Jang · eric@jangq.ai
• Base model © MiniMax, used under the MiniMax-M3 license.