Large Language Models

MiniMax M3

The first open-weight model to combine frontier coding, a 1M-token context and native multimodality — a 428B MoE (22B active) built on sparse attention.

MiniMax M3, released June 1 2026 by Shanghai-based MiniMax, makes a specific claim: it’s the first open-weight model to combine frontier-level coding, a 1M-token context and native multimodality (text, image, video) in a single system. It’s a 428B-parameter mixture-of-experts with ~22B active per token, trained multimodally from step zero across ~100 trillion interleaved tokens.

The enabling trick is MiniMax Sparse Attention (MSA), which pre-filters context blocks instead of attending to everything — cutting per-token compute to roughly 1/20th of M2 at 1M context, with far faster prefill and decoding. That’s what turns a 1M window into a production feature rather than a spec sheet number. On SWE-Bench Pro it scores ~59%, ahead of GPT-5.5 and Gemini 3.1 Pro on that benchmark, at a modest ~$0.60 / $2.40 per 1M tokens.

Two honest caveats. The license is the MiniMax Community License — weights are downloadable on Hugging Face, but it’s source-available rather than OSI-open, with commercial-use restrictions to check before shipping. And the 1M context is “up to,” with 512K guaranteed; plan around 512K, and expect a multi-GPU node for self-hosting.

Best for

  • Long-context agentic coding and 8h+ sessions
  • Long-video and multi-document understanding
  • Self-hosted multimodal deployments

Pros & cons

Strengths

  • Frontier coding + 1M context + native multimodality in one open model
  • SWE-bench Pro ~59, ahead of GPT-5.5 and Gemini 3.1 Pro on that test
  • Sparse Attention (MSA) makes long context cheap (~1/20th M2's per-token compute)

Limitations

  • Community license restricts commercial use (not OSI-open)
  • 428B MoE needs a multi-GPU node; 1M is "up to," 512K guaranteed

Sources

Last updated: 2026-06-18 · Specs and pricing change fast — verify on the vendor's site before relying on them.