Superseded (verified 2026-10-01)
This note covers the M2 line (M2.1, Dec 2025). Newer open releases: M2.5, M2.7 (2026-03-18) and the current flagship minimax-m3 (2026-05-31/06-02).
MiniMax M2.1
MiniMax’s coding- and agent-focused open-weight MoE model.
Lineage (then → 2026 now)
| Model | Date | Notes |
|---|---|---|
| M2 | HF repo last updated 2025-12-23 (original date not verified) | 229B total (HF), predecessor |
| M2.1 | 2025-12-23 (MiniMax news) | this note; SWE-bench Verified 74.0 |
| M2.5 | HF repo updated 2026-03-10 | 229B (HF) |
| M2.7 | released 2026-03-18 (MiniMax news page, re-fetched 2026-10-02); weights 2026-04-12 (press, not re-checked) | “first model to participate in its own evolution”; SWE-Pro 56.22 matching GPT-5.3-Codex, GDPval-AA 1495 ELO, native Agent Teams (all vendor); via MiniMax Agent and API |
| M3 | June 2026 (arXiv paper “MiniMax Sparse Attention” submitted 2026-06-11; earlier-reported 05-31/06-02 dates unverified), see minimax-m3 | ~428B total / ~23B active, 1M context, native multimodal, minimax-community licence; SWE-bench Verified 80.5 (card) |
Specifications of M2.1 (Hugging Face card)
- Parameters: 229B total (HF card). The “~10B active” figure from earlier versions is not on the card and was removed.
- Licence: Modified MIT
- Release: 2025-12-23 (MiniMax announcement). Text-generation only.
- Serving: Transformers, vLLM, SGLang, KTransformers, MLX-LM
- Sampling advice: temperature 1.0, top_p 0.95, top_k 40
Benchmarks (card, vendor-reported)
SWE-bench Verified 74.0 (M2: 69.4); SWE-bench Multilingual 72.5 (Opus 4.5: 77.5 per the card); VIBE full-stack 88.6; MMLU-Pro 88 (card). Multi-SWE-Bench 49.4 and GPQA-Diamond 83.0 from the earlier read were not on the re-read card and are dropped.
Plans
The MiniMax pay-as-you-go page lists M3 with usage-based token pricing (up to 512K context, higher rate above; a priority tier at a higher rate) with a “permanent 50% discount” shown, and older models M2.7 and M2.5 undiscounted. M2.1-specific pricing is not shown (not a key claim). Current prices: see https://platform.minimax.io/docs/guides/pricing-paygo
What changed in 2026
Four newer text models in nine months (M2.5, M2.7, M3, plus H3 video and Music 3.0). Use M3 for new work; M2.x remain open-weight for local use.
Open items
- M2.1 per-model API price, the M2 original release date and the M2.7 weights date (04-12) are not on primary pages; none is a key claim.
See also
- MiniMax M1, MiniMax-Text-01, minimax-m3
Sources
- https://huggingface.co/MiniMaxAI/MiniMax-M2.1 (accessed 2026-10-01, re-fetched 2026-10-02)
- https://arxiv.org/abs/2606.13392 (accessed 2026-10-02)
- https://huggingface.co/MiniMaxAI (org page, sizes and update dates; accessed 2026-10-01)
- https://www.minimax.io/news/minimax-m21 (accessed 2026-10-01)
- https://www.minimax.io/news/minimax-m27-en (accessed 2026-10-01)
- https://huggingface.co/MiniMaxAI/MiniMax-M3 (accessed 2026-10-01)
- https://platform.minimax.io/docs/guides/pricing-paygo (accessed 2026-10-01)
- https://pandaily.com/mini-max-open-sources-m2-7-model-with-broad-hardware-ecosystem-support (search result, weights date; accessed 2026-10-01)