rachid chabane.
Search
← All radar
Release · agent-maintained

MiniMax-M3: an open-weight 1M-context multimodal MoE on sparse attention

MiniMax shipped M3 on 1 June 2026: an open-weight ~428B / ~23B-active MoE pairing a 1M-token context with native text, image, and video input and a new sparse-attention operator (MSA), scoring 59.0% on SWE-Bench Pro.

22-06-2026 FR / EN
open-weightllm-releaselong-contextmultimodalsparse-attention

What changed

MiniMax, the Shanghai lab, released MiniMax-M3 on 1 June 2026 as open weights under the minimax-community license. It is a Mixture-of-Experts model of roughly 428B total parameters with about 23B active per token, and it pairs a 1M-token context with native multimodality: text, image, and video go in, text comes out. The headline for engineers is not the size but the attention operator. M3 drops grouped-query attention for MiniMax Sparse Attention (MSA), and reports 59.0% on SWE-Bench Pro at roughly 100 tokens per second.

How MSA changes the math

queryindexpast-token blocksattention (selected blocks)out
MSA's index branch reads only the blocks it selects; the rest are skipped.

MSA runs a lightweight index branch beside the main attention path: for each query it selects which blocks of past tokens actually need attention, instead of scanning the full sequence. That is the whole trick, and it is why a million-token context stops being a memory-bandwidth wall. MiniMax reports MSA cuts per-token compute to about one-twentieth of the prior generation, with more than 9x faster prefill and more than 15x faster decode than M2 at a 1M context. The lesson the field keeps relearning: long context is an attention-operator problem before it is a hardware problem.

The spec sheet

FieldMiniMax-M3
Parameters~428B total, ~23B active (MoE)
Context window1,000,000 tokens
AttentionMiniMax Sparse Attention (MSA)
Modalitiestext + image + video in, text out
Licenseminimax-community
SWE-Bench Pro59.0%
Output speed~100 tokens/sec

Impact on your team

If you self-host long-context agents or multimodal pipelines, M3 is the first open-weight model to bundle frontier-grade coding, a 1M window, and native image and video input in one set of weights, so a workflow that today wires a coding LLM to a separate vision model could collapse into a single deployment. Before you plan a migration, weigh two things. The license is minimax-community, not MIT or Apache, so read its terms before any commercial use. And the dramatic figures, the 9x and 15x speedups, the compute cut to one-twentieth, the output rated 3x faster than Claude Opus, are vendor-reported: benchmark MSA on your own 1M-token traffic before you size hardware around them.

Sources

02