YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
minimax-m3: 2 representative decoder layers
Source: MiniMaxAI/MiniMax-M3.
Original text dimensions, context/RoPE and precision settings are retained. Text-only benchmark; vision is skipped. No weights are included. Use native SGLang with --load-format dummy.
| Layer | Source layer (zero-based) | Attention | FFN | PLE |
|---|---|---|---|---|
| 0 | 0 | GQA | dense | no |
| 1 | 3 | block-sparse GQA | MoE | no |
Separate MTP modules retained: 1. Ordinary one_batch decoding does not execute MTP.
Run bash minimax-m3-2layer.run from the checkout root on a CUDA machine. See docs/frontier-layer-types.md for native SGLang requirements and layer reasoning. The manifest records provenance; it is not a runtime input.
- Downloads last month
- 24
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support