MiniMax M3 is HERE! (Real Tests and Review)

MiniMax M3 is the first open-weight model to ship 1M-token context, agentic coding, and native multimodal training in a single release — but open weights drop in 10 days, so today's tests are mostly the desktop app.

What the benchmarks actually say

On SWE Bench Pro, MiniMax M3 scores 59%, sitting 0.4 points above GPT 5.5 but roughly 10% behind Claude Opus 4.8's 69.2%. Treat SWE Bench Pro as the real coding bar — the creator calls it "the most reliable indicator of how strong the model is for coding" and explicitly notes it "is not cherry-picked." On OmniDock document understanding, M3 lands above Gemini 3.1 Pro.

Where M3 actually wins: speed and cost

The engine is MiniMax Sparse Attention (MSA), which drops per-token compute to "roughly about 1/20 of what the previous generation needed." At 1M tokens, M3 pre-fills 9.7× faster and decodes 15.6× faster than M2.7, measured specifically at the "worst-case long end, exactly where dense models typically crawl." API pricing is $0.30 per 1M input tokens and $1.20 per 1M output tokens.

How to actually use M3 today

Download the desktop app called MiniMax Code. Treat it as a "cloud co-work" with built-in web search for research, presentations, and landing pages — not as a coding harness. Kilo Code and similar harnesses still show M2.7 on the token plan; M3 is on Kilo's pay-per-use gateway only, with a 5-hour token reset on MiniMax's side.

The catches

The first 3D Chinese-architecture test rendered the model and dashboard but the roof was "still messy put together," and labels shift as the camera moves. Video generation is gated behind the Max plan (3 per day; Ultra gives 5) — the Plus plan has no video. Context inside the MiniMax token plan is still capped at 200K, and M3 multimodal only unlocks once open weights land.

What to watch

Open-weight release in "the next 10 days" will unlock Kilo Code and Hermes Agent integrations — that's when to re-test properly.