MiniMax M3
Listed as the newest flagship model of MiniMax on its Hugging Face page (checked 2026-09-30). Described by the model card as a native multimodal sparse-attention MoE with weights on Hugging Face.
Specifications (Hugging Face card)
- Parameters: ~428B total / ~23B active; an earlier version of this note listed an MXFP8 440B build; the card re-read on 2026-10-05 does not mention one (it only links MXFP4/MXFP8 quantisation guides), so it is dropped
- Context: 1M tokens
- Modalities: text, image, video (mixed-modality training “from the very first step”)
- Attention: MiniMax Sparse Attention; card claims 9x prefill and 15x decode speedups vs M2 at 1M context
- Reasoning modes: enabled, adaptive, disabled
- Licence: MiniMax Community License (LICENSE file read 2026-10-05): non-commercial use unrestricted; commercial use allowed if “Built with MiniMax M3” is displayed; products with revenue above USD 20M a year need MiniMax’s prior written authorisation, below that a one-time notice to MiniMax; usage restrictions (illegal content, military use, harm to minors, disinformation, hate speech). Read the LICENSE itself before commercial use.
- Release: June 2026 (the card’s arXiv paper 2606.13392, “MiniMax Sparse Attention”, was submitted 2026-06-11; earlier audit text said 2026-06-01 / 05-31 announcement). Exact release day unverified. Repo last updated Jul 23.
- Access: MiniMax Agent, API platform, SGLang / vLLM / Transformers;
thinkingparameter selects the reasoning mode; recommended temperature 1.0, top_p 0.95 (card, 2026-10-02). - Pricing (MiniMax pay-as-you-go page, 2026-10-02): usage-based per-token pricing up to 512K context (shown with a “permanent 50% discount”), double the rate above 512K; a priority tier at a higher rate. Older M2.7 / M2.5 are undiscounted. Current prices: see https://platform.minimax.io/docs/guides/pricing-paygo
Benchmarks (vendor-reported)
SWE-bench Verified 80.5%; SWE-bench Pro 59%; MMMU Pro 78.1%; Video-MME v2 85.4%; Apex Agents 27.7%.
Lineage
minimax-text-01 → minimax-m1 → minimax-m2 (M2.1, M2.5, M2.7) → M3.
Open items
- Exact release day: not on a primary page (minimax.io news pages 404 or undated); the arXiv paper (MiniMax Sparse Attention, 2606.13392, v1 2026-06-11) says MiniMax-M3 has been released, so release was June 2026 or earlier. Benchmarks are vendor-reported.
Sources
- https://huggingface.co/MiniMaxAI/MiniMax-M3 (accessed 2026-09-30, re-fetched 2026-10-02)
- https://huggingface.co/MiniMaxAI/MiniMax-M3/blob/main/LICENSE (2026-10-05)
- https://platform.minimax.io/docs/guides/pricing-paygo and https://arxiv.org/abs/2606.13392 (2026-10-02)
- https://huggingface.co/MiniMaxAI (accessed 2026-09-30)
- https://www.minimax.io/news (model list, no dates, accessed 2026-09-30)