MiniMax-M3
Released Jun 1, 2026 · knowledge cutoff unpublished
- Status
- Unknown
- Location
- China
- Modality
- Multimodal
- Context window
- 1M
- Max output
- —
- Speed
- 103 tok/s · 1.49s to first answer token
- Price ($/MTok in / out)
- $0.30 / $1.20
- Cost per task
- —
- Open weights
- Yes
Benchmarks2 of 8 reported
SWE-bench Verified
Terminal-Bench 2.1
MMLU-Pro
GPQA Diamond
AIME (latest)
Humanity's Last Exam
LMArena Elo
ARC-AGI-2
Strengths & weaknesses
Strengths
- First open-weight model from any lab to combine a 1M-token context window, frontier-tier coding, and native image/video understanding in one checkpoint
- New MiniMax Sparse Attention architecture gives large prefill/decode speedups over M2 at long context, so the 1M window is reportedly usable rather than nominal
- Can drive a desktop computer-use agent (MiniMax Code) directly, not just chat
Weaknesses
- License tightened from M2's largely unrestricted terms to a revenue-gated community license requiring authorization above $20M annual revenue
- Multimodal input is new to this generation, so image/video handling is comparatively less battle-tested than its text/coding path
- Early hands-on reception is cautiously optimistic rather than settled — reliability on production agent loops still needs workload-specific testing
Officially announced 2026-06-01 (minimax.io/blog/minimax-m3); some third-party trackers (e.g. OpenRouter) list 2026-05-31, likely the weights-availability date. Total/active parameters ~428B/23B per the Hugging Face model card. SWE-bench Verified 80.5% and Terminal-Bench 2.1 66.0% are corroborated across the Hugging Face README and MiniMax's own blog. GPQA Diamond, AIME, HLE, MMLU-Pro and ARC-AGI-2 were not extractable — MiniMax's benchmark table at launch is an embedded image rather than machine-readable text, and no independent source reproduced these specific figures. Licensed under the bespoke 'MiniMax Community License': free for non-commercial use; commercial use under $20M annual revenue requires only a notice email and a 'Built with MiniMax' attribution label, above that threshold prior written authorization is required. Third-party framing calls M3 a 'successor to M2.7' (not tracked here), but MiniMax's own material explicitly keeps M2.5 available as a cheaper option for some workloads rather than deprecating it, so predecessorId is left null. Pricing is the standard tier for inputs up to 512K tokens; above that the rate doubles to $0.60/$2.40 per MTok.
For developers
API model strings
- minimax
MiniMax-M3
Licence
MiniMax Community License · Restricted · commercial use permitted
Retirement
No retirement announced
Lineage
No recorded predecessor