MiniMax M3

$0.30 per 1M input tokens and $1.20 per 1M output tokens; cached input $0.06 per 1M; cheapest via CoreWeave at $0.23 / $0.96. A typical benchmark task costs about $0.51 at default effort.

There is no free API; the cheapest paid price is $0.23 per 1M input tokens. The weights are open, so you can run it yourself.

#49 of 334 in Chat & Knowledge, #49 of 132 in Coding, #58 of 133 in Agents & Tool Use, #62 of 195 in Writing & Language, #29 of 124 in Vision & Documents. Strongest at: Tool calling, Reasoning, Science, Long context, Knowledge, Instructions, Vision.

On Arena Text · creative writing, MiniMax M3 scores 1,405 vs 1,364 — higher.

Through the API of CoreWeave, DeepInfra, Featherless, Mara, Minimax and 6 more.

524K tokens, with up to 512K output tokens.

Released 2026-06-01. It is the current version in its line.

Yes, the weights are open (minimax-community licence).

The weights are open and can be self-hosted; hardware needs are not confirmed yet.

It accepts text and images, video as input.

MiniMax M3 is an open-weights language model from MiniMax, released Jun 2026, the successor to MiniMax M2.7.

GLM 5.3 Flash (cheaper), GPT-5.6 Terra (better), GLM 5.2 (open weights).