MiniMax-M3

How it is built

MiniMax-M3 is a mixture of experts, and its card gives the size only roughly: about 428B parameters, about 23B of them active for each token.¹¹1MiniMax, MiniMax-M3 model card on Hugging Face: "~428B parameters and ~23B activated parameters"; mixed-modality training "from the very first step"; 9x prefill and 15x decode against M2 at 1M context, per-token compute cut to 1/20; thinking modes enabled, adaptive, disabled. https://huggingface.co/MiniMaxAI/MiniMax-M3. The MXFP8 repository is https://huggingface.co/MiniMaxAI/MiniMax-M3-MXFP8

Three dense layers come first. In each of the other 57, the router keeps just 4 of 128 routed experts, beside 1 shared expert.²²2MiniMax-M3 config.json, text_config: num_hidden_layers 60, moe_layer_freq 0 for the first 3 layers and 1 for the other 57, num_local_experts 128, n_shared_experts 1, num_experts_per_tok 4, scoring_func sigmoid, use_routing_bias true, num_attention_heads 64, num_key_value_heads 4, sparse attention on the same 57 layers with sparse_block_size 128 and sparse_topk_blocks 16; vision_config num_hidden_layers 32. https://huggingface.co/MiniMaxAI/MiniMax-M3/blob/main/config.json. Arithmetic: 60 minus 3 = 57. MiniMax-M2.7 sends a token to 8 of 256 and has no shared expert.³³3MiniMax-M2.7 config.json: num_local_experts 256, shared_intermediate_size 0, num_experts_per_tok 8. https://huggingface.co/MiniMaxAI/MiniMax-M2.7/blob/main/config.json

Every layer uses grouped query attention, with 64 query heads over 4 key-value heads.²² The first 3 layers attend in full. The other 57 use MiniMax Sparse Attention.²²

An index branch scores blocks of earlier tokens and picks a few for each head group.⁴⁴4MiniMax, "MiniMax Sparse Attention", arXiv 2606.13392, 11 June 2026: blockwise sparse attention built on GQA, index branch with per-group top-k block selection; on a 109B test model, 28.4x less attention compute per token at 1M context, on par with GQA, 14.2x prefill and 7.6x decode wall-clock speedups on H800. https://arxiv.org/abs/2606.13392 The main branch then runs exact attention over those blocks only. In M3 a block is 128 tokens, and 16 blocks are kept.²²

Images and video entered training alongside text "from the very first step".¹¹

What is new

Against M2 at a 1M context, MiniMax reports prefill 9 times faster and decoding 15 times faster.¹¹ Compute per token falls to a twentieth.¹¹

The attention paper tested the method on a separate 109B model. At 1M tokens it cut attention compute per token 28.4 times.⁴⁴ On H800 GPUs, prefill ran 14.2 times faster and decoding 7.6 times faster.⁴⁴ Those numbers describe the test model, not M3.

The M2 series accepts only text and tool calls. M3 adds images and video.⁵⁵5MiniMax, Anthropic-compatible API reference: M3 supports text, image and video blocks; "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

MiniMax puts M3 at 83.5 on BrowseComp, above Claude Opus 4.7 at 79.3, and publishes no settings for the run.⁶⁶6MiniMax, MiniMax M3 model page: context up to 1M with a guaranteed minimum of 512K; BrowseComp 83.5 against 79.3 for Opus 4.7, no run settings given. https://www.minimax.io/models/text/m3

Running it

MiniMax-M3 opened on the API on 1 June 2026.⁷⁷7MiniMax, API release notes, MiniMax M3 entry dated 1 June 2026. https://platform.minimax.io/docs/release-notes/models The context runs to 1M tokens, with at least 512K guaranteed.⁶⁶⁸⁸8MiniMax, text generation guide, model table: MiniMax-M3 context window 1,000,000. https://platform.minimax.io/docs/guides/text-generation Reasoning can be turned off or left to the model to decide.¹¹

MiniMax marks its M3 prices "Permanent 50% off". With the discount, prompts up to 512K tokens cost $0.30 per million input tokens and $1.20 per million output tokens.⁹⁹9MiniMax, pay-as-you-go pricing, USD per million tokens, read 29 September 2026, marked "Permanent 50% off". Up to 512K input: 0.30 input, 1.20 output, 0.06 cache read, against list prices of 0.60, 2.40 and 0.12. Above 512K: 0.60, 2.40 and 0.12, against 1.20, 4.80 and 0.24. https://platform.minimax.io/docs/guides/pricing-paygo. Arithmetic: 0.60 / 0.30 = 2; 2.40 / 1.20 = 2. Longer prompts cost twice that.⁹⁹

The BF16 weights on Hugging Face fill 854.2 GB. An MXFP8 copy in a second repository takes 443.7 GB.¹¹¹⁰¹⁰10Sum of the .safetensors file sizes listed by Hugging Face, read 29 September 2026: MiniMaxAI/MiniMax-M3, 59 files, 854.2 GB; MiniMaxAI/MiniMax-M3-MXFP8, 31 files, 443.7 GB. https://huggingface.co/MiniMaxAI/MiniMax-M3/tree/main and https://huggingface.co/MiniMaxAI/MiniMax-M3-MXFP8/tree/main

Non-commercial use is free under the MiniMax Community License.¹¹¹¹11MiniMax Community License, clauses 2 and 3 and the appendix of prohibited uses. https://huggingface.co/MiniMaxAI/MiniMax-M3/blob/main/LICENSE A commercial product must carry a visible "Built with MiniMax M3" credit, and its maker must send MiniMax a one-time notice by email. Above US$20 million of yearly revenue, written authorization comes first. Military use is prohibited.¹¹¹¹