MiniMax slips M3.1-Flash into MiniMax Code with five effort tiers

The fast tier defaults to max reasoning but stays locked inside the app — no API, weights or benchmarks yet, while a free 1.6T LongCat runs on OpenCode.

Nowline SEP 28 2:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • A faster coding tier with five reasoning dials

    MiniMax dropped M3.1-Flash-Preview into MiniMax Code on Sept 27 — a speed-tuned sibling to M3 with five effort levels (low, medium, high, xhigh, max) and a default of max. Sparse attention plus 4-bit experts and cache aim it at routine bug fixes and reviews where you want low latency, not long-horizon agent runs.

  • The catch: it's boxed inside one app

    There's no public API, no published pricing, no benchmarks, and the ~250GB weights sit in a gated Hugging Face repo. Until MiniMax opens an endpoint you can't wire M3.1-Flash into your own stack — it lives only in MiniMax Code's model picker, so treat it as a preview to try, not a dependency to ship on.

  • Meanwhile, a 1.6T model you can actually call — free

    LongCat-2.5-Preview is free on OpenCode for two weeks through an OpenAI-compatible endpoint (model id longcat-2.5-preview-free), with a 1M-token context, ~48B active params on a 1.6T MoE, native multimodality, and zero data retention. Point an existing client at it and stress-test long-context or vision work at no cost this weekend.

  • But treat LongCat as a tire-kick, not a default

    It's a 'Preview' with no published weights and no benchmarks — vendor, independent, or leaderboard. One early tester called the prior LongCat 2.0 'completely hopeless' on small tasks, so run your own evals before routing anything real through it, especially since the free window has no confirmed end date.