MiniMax-M3

2026-06-01 · Publisher : MiniMax · Domain : Artificial Intelligence

MiniMax releases M3, a multimodal MoE of about 428 billion parameters (23 billion active) with one million tokens of context thanks to sparse attention, and weights under a community license.

### Context The M2 series (October 2025 to March 2026) was limited to about 200,000 tokens and to text, and its license tightened with M2.7. In June 2026, the closed references cited by MiniMax are Claude Opus 4.7 and GPT-5.5. The vendor claims the first open model to combine top-level coding, a one-million-token context and native multimodality. ### What the model brings - Native multimodality: the data pipeline was rebuilt with multimodal training from the first step, for text, image and video. - MiniMax Sparse Attention (MSA): block-wise sparse attention, with an indexing branch that selects a subset of key-value blocks; the model card claims 9 times faster prefill and 15 times faster decoding than M2 at 1 million tokens. - Context: up to 1 million tokens via the API, with 512,000 tokens guaranteed at minimum according to the product page. - Three reasoning modes: on, adaptive or off. - Long autonomous tasks: replication of an ICLR paper in nearly 12 hours and optimization of an FP8 GEMM kernel with a 9.4-fold gain, according to the vendor. ### Architecture and specifications - Parameters: about 428 billion in total, about 23 billion active; 60 layers, 128 routed experts of which 4 active per token (config.json). - Context: 1,048,576 positions (config.json). Modalities: text, image and video as input. - Training data and knowledge cutoff: not disclosed. - License: minimax-community (copyright 2026). Non-commercial use free. Commercial use: one-time notification to api@minimax.io under $20 million of annual revenue, prior written authorization above that, and a “Built with MiniMax M3” notice displayed. A prohibited-use policy applies (illegal content, military use, minors, disinformation, discrimination). - Weights: https://huggingface.co/MiniMaxAI/MiniMax-M3 (MXFP8 version also published), with SGLang, vLLM, Transformers and KTransformers. ### Cost and access | Item | Value | |---|---| | Standard, input (512,000 tokens or fewer) | $0.30 per million tokens | | Standard, output (512,000 tokens or fewer) | $1.20 per million tokens | | Standard, beyond 512,000 tokens | $0.60 input, $2.40 output | | Cache read | $0.06 ($0.12 beyond 512,000) | | Priority tier | 1.5 times the standard price | | Self-hosting | hardware not disclosed in the sources consulted | Pricing from the official pricing page, which shows a permanent 50% discount. ### Benchmarks Figures published by the vendor (product page and model card). | Test | MiniMax-M3 | Closed model cited | |---|---|---| | BrowseComp | 83.5 | Claude Opus 4.7: 79.3 | | PostTrainBench | 37.1 (3rd rank) | Opus 4.7: 42.4; GPT-5.5: 39.3 | | SWE-bench Verified | 80.5 | not compared in the primary sources read | | SWE-Bench Pro | 59.0 | not compared in the primary sources read | | MMMU-Pro | 78.1 | not compared in the primary sources read | Source: MiniMax M3 product page and model card (.eval_results file), vendor measurements. The model card's full comparison table is an image that was not used. ### Versus closed models At the time of its release, the vendor positions M3 against Claude Opus 4.7 and GPT-5.5. On BrowseComp, it claims 83.5 versus 79.3 for Opus 4.7. On PostTrainBench (autonomous model training), M3 scores 37.1, third place behind Opus 4.7 (42.4) and GPT-5.5 (39.3), and far ahead of the other models according to the vendor. The standard price of $0.30 input and $1.20 output is not compared figure by figure with closed pricing in the primary sources read. For the open-or-closed choice, the license allows self-hosting and commercial use with notification under $20 million of revenue, which is more permissive than M2.7 and less so than M1's Apache 2.0; beyond the threshold, an authorization is required. ### Key takeaways M3 is MiniMax's multimodal and long-context milestone, with a technical report on sparse attention. Numerical comparisons against closed models are limited to the vendor's figures and a few tests, which calls for independent measurements before a production choice.

Tags : LLM, Open Source, MiniMax, MoE, Multimodal, Contexte long, minimax-community