MiniMax-M2.7

How it is built

MiniMax-M2.7 reuses the dimensions of MiniMax-M2. The two configs match in every model dimension and differ only in the maximum number of positions, 204,800 against 196,608.¹¹1MiniMax-M2.7 config.json: num_hidden_layers 62, num_local_experts 256, shared_intermediate_size 0, num_experts_per_tok 8, scoring_func sigmoid, use_routing_bias true, attn_type_list the same for all 62 layers, num_attention_heads 48, num_key_value_heads 8, num_mtp_modules 3, max_position_embeddings 204800. https://huggingface.co/MiniMaxAI/MiniMax-M2.7/blob/main/config.json ²²2MiniMax-M2 config.json, compared field by field with M2.7's: hidden size, layers, heads, experts, experts per token and intermediate sizes are identical; max_position_embeddings is 196608. https://huggingface.co/MiniMaxAI/MiniMax-M2/blob/main/config.json

There are no dense layers. All 62 hold 256 routed experts and no shared expert, and a sigmoid router with a routing bias picks 8 of them per token.¹¹

Attention is the same in every layer, with 48 query heads sharing 8 key and value heads.¹¹ The config also defines three multi-token prediction modules.¹¹

MiniMax has published no parameter count for M2.7. For M2 it gave 230B in total and 10B active.³³3MiniMax, MiniMax-M2 model card: "230 billion total parameters with 10 billion active parameters". https://huggingface.co/MiniMaxAI/MiniMax-M2 Hugging Face counts 228.7B parameters in the M2.7 weight files, which rounds to 229B. It counts exactly the same number for M2.⁴⁴4Hugging Face model API, safetensors totals read 29 September 2026: MiniMax-M2.7 228,689,764,864 parameters, 227,410,968,576 of them in F8_E4M3; MiniMax-M2 228,689,764,864. This is Hugging Face's count of the published files, not a figure stated by MiniMax. The .safetensors files of MiniMax-M2.7 add up to 230.1 GB. https://huggingface.co/api/models/MiniMaxAI/MiniMax-M2.7 and https://huggingface.co/api/models/MiniMaxAI/MiniMax-M2. Arithmetic: 227.41 / 228.69 = 99.4% of parameters in FP8; 228.69B rounds to 229B. MiniMax states no active count for M2.7.

What is new

MiniMax calls M2.7 its "first model deeply participating in its own evolution".⁵⁵5MiniMax, MiniMax-M2.7 model card on Hugging Face: self-evolution, 100+ rounds, 30% improvement; SWE-Pro 56.22%; Terminal Bench 2 57.0%; MLE Bench Lite, 22 competitions, 66.6% medal rate. https://huggingface.co/MiniMaxAI/MiniMax-M2.7 An internal version optimized its own programming scaffold over more than 100 rounds, and MiniMax reports a 30% gain from that loop.⁵⁵

The card reports 56.22% on SWE-Pro and 57.0% on Terminal Bench 2. On MLE Bench Lite, 22 machine learning competitions, it wins a medal in 66.6% of them.⁵⁵ None of these scores comes with run settings.

The licence changed. M2 and M2.5 were released under a modified MIT licence.⁶⁶6MiniMax-M2 and MiniMax-M2.5 model cards, licence field "modified-mit". https://huggingface.co/MiniMaxAI/MiniMax-M2 and https://huggingface.co/MiniMaxAI/MiniMax-M2.5 M2.7 requires MiniMax's written authorization for any commercial use.⁷⁷7MiniMax-M2.7 LICENSE, "NON-COMMERCIAL LICENSE", clauses 2, 3 and 5 and the appendix of prohibited uses. https://huggingface.co/MiniMaxAI/MiniMax-M2.7/blob/main/LICENSE

Running it

MiniMax-M2.7 reached the API on 18 March 2026.⁸⁸8MiniMax, API release notes, MiniMax M2.7 entry dated 18 March 2026. https://platform.minimax.io/docs/release-notes/models A second id, MiniMax-M2.7-highspeed, promises the same performance at a higher speed. MiniMax quotes about 100 tokens per second for it and about 60 for the standard id.⁹⁹9MiniMax, text generation guide, model table: MiniMax-M2.7 and MiniMax-M2.7-highspeed, context window 204,800, about 60 and about 100 tokens per second. https://platform.minimax.io/docs/guides/text-generation

The context is 204,800 tokens.⁹⁹ Only text and tool calls go in, with no images.¹⁰¹⁰10MiniMax, Anthropic-compatible API reference: "The M2.7, M2.5, M2.1, and M2 series support text and tool-call content blocks only". https://platform.minimax.io/docs/api-reference/text-anthropic-api The standard id costs $0.30 per million input tokens and $1.20 per million output tokens, and the fast one twice that.¹¹¹¹11MiniMax, pay-as-you-go pricing, USD per million tokens, read 29 September 2026: MiniMax-M2.7 0.30 input, 1.20 output, 0.06 cache read, 0.375 cache write; MiniMax-M2.7-highspeed 0.60, 2.40, 0.06, 0.375. https://platform.minimax.io/docs/guides/pricing-paygo. Arithmetic: 0.60 / 0.30 = 2; 2.40 / 1.20 = 2.

The weights are almost entirely FP8, 99.4% of the parameters, in 230.1 GB of files.⁴⁴

Personal use is free, self-hosted deployment for coding and research included, and so is non-profit and academic research.⁷⁷ Commercial use of any kind needs MiniMax's prior written consent, requested by email. A commercial product must also show "Built with MiniMax M2.7". Military use is prohibited.⁷⁷