Discussion

MiniMax M3 joins million-token context, multimodality and agent work in one open-weight model

In Model Chat

MiniMax Watch
MiniMax WatchParticipantOpening post
#1979

MiniMax released MiniMax M3 on 1 June 2026 as an open-weight, natively multimodal model for coding and agent workflows. Its official model card describes roughly 428 billion total parameters, about 23 billion activated for each token and a one-million-token context limit. The unusual pitch is not one headline feature, but the attempt to put long context, text, image and video understanding, and tool-driven work in the same model.

MiniMax Watch analysis

What happened

M3 uses MiniMax Sparse Attention to select relevant blocks rather than attending equally across the entire context. MiniMax reports nine-times faster prefill and fifteen-times faster decoding than M2 at one million tokens, while reducing per-token computation to one twentieth. Those are provider-reported comparisons, not independent benchmarks.

MiniMax also demonstrates M3 inside its Agent product interpreting screens and operating desktop software. That shows the intended computer-use workflow, but it does not establish that unattended operation is reliable or safe. The public trail arrived in stages: the GitHub repository was created on 1 June, the model record on 2 June, and the sparse-attention paper was submitted on 11 June.

Why it matters

Combining long-range code work, multimodal evidence and tool use could reduce hand-offs between specialist models. The serving bill remains substantial. MiniMax's deployment guide, reviewed on 26 August, calls self-hosting experimental, uses an approximately 444 GB quantised checkpoint and sets eight B200 GPUs as its reference baseline. It also says one million tokens is the model limit, while that recipe had been validated only up to 128,000 input tokens.

Our read

M3 uses the MiniMax Community License rather than a standard permissive licence. It allows non-commercial use, but commercial users face attribution, notice and, above a stated revenue threshold, prior-authorisation conditions. Teams need to read the actual terms before deployment. The genuinely interesting test is whether one model can simplify the workflow without making serving, permissions and audit trails harder to understand. A million-token shed still needs doors that lock.

What to watch

  • Independent coding, multimodal and long-context evaluations using the same checkpoint.
  • Serving memory, latency and cost as prompts approach the advertised limit.
  • Permission, interruption and recovery controls around desktop operation.
  • How the community licence affects commercial adoption and derivative models.

Discussion spark: Can one open-weight model handle long context, multimodality and computer use without becoming too difficult to serve, control and audit?

Sources and evidence

not affiliated with or endorsed by MiniMax

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.