MiniMax-M3
A 428B open MoE with real 1M context and native image/video input at $0.30/$1.20. Vendor coding claims (80.5% SWE-bench Verified) outrun its AA score of 29. Good value as a worker agent; weak as the brain of the operation.
Key facts
- Score
- 64 / 100, rank #18Evals56Price67People55Fit100
- Price
- $0.30 in · $1.20 out per 1M tokens (MiniMax-M3 - Artificial Analysis)
- Blended price
- $0.52 per 1M tokens (3 input : 1 output)
- Context window
- 1M
- Open weights
- Yes
- Released
- Jun 1, 2026
- Best for
- Long-context document and video analysis, Cheap worker agents under a smarter planner, Multimodal RAG
MiniMax-M3 costs $0.30 per million input tokens and $1.20 per million output tokens on MiniMax's list price (MiniMax-M3 - Artificial Analysis, read Oct 11, 2026).
Independent evals
| Benchmark | Result | Who ran it | Read |
|---|---|---|---|
| Artificial Analysis Intelligence Index | 29 | MiniMax-M3 - Artificial Analysis | Oct 10, 2026 |
| Arena Text score | 1440 (rank 96) | Text Arena Leaderboard | Oct 10, 2026 |
| SWE-bench Verified (vendor) | 80.5% | MiniMaxAI/MiniMax-M3 - Hugging Face | Oct 10, 2026 |
- Artificial Analysis output speed (tokens/s): 108.2 (MiniMax-M3 - Artificial Analysis)
What practitioners say
Quieter launch than GLM or MiMo. In the HN thread on a MiniMax M3 vs GLM 5.2 codegen comparison, the consensus was that M3 'works best as a worker agent' ('create plans with a smarter model, then use Minimax to execute'), that 'their token plan is very cheap', but that as a top-level agent it 'gets way too confused' next to GLM-5.2 or GPT-5.4 mini. The linked comparison concluded GLM 5.2 wins from-scratch builds while M3 wins on worker economics.
Caveats
- Headline SWE-bench numbers were run by MiniMax with a Claude Code harness; independent replication lags.
- HN users found it 'gets way too confused' as a top-level agent compared with GPT-5.4 mini or GLM-5.2.
- Custom 'minimax-community' licence, not MIT or Apache; check restrictions.
Questions people ask
How much does MiniMax-M3 cost?
MiniMax-M3 costs $0.30 per million input tokens and $1.20 per million output tokens on MiniMax's list price (MiniMax-M3 - Artificial Analysis, read Oct 11, 2026).
How good is MiniMax-M3?
MiniMax-M3 ranks #18 of 26 on the Models at Work leaderboard with a score of 64 out of 100 (October 2026 edition). A 428B open MoE with real 1M context and native image/video input at $0.30/$1.20. Vendor coding claims (80.5% SWE-bench Verified) outrun its AA score of 29. Good value as a worker agent; weak as the brain of the operation.
What is MiniMax-M3 best for?
MiniMax-M3 is best for long-context document and video analysis, cheap worker agents under a smarter planner, multimodal rag.
What are MiniMax-M3's benchmark scores?
Artificial Analysis Intelligence Index: 29 (MiniMax-M3 - Artificial Analysis); Arena Text score: 1440 (rank 96) (Text Arena Leaderboard); SWE-bench Verified (vendor): 80.5% (MiniMaxAI/MiniMax-M3 - Hugging Face).
What is MiniMax-M3's context window?
MiniMax-M3 has a 1M token context window according to MiniMax.
Is MiniMax-M3 open weights?
Yes. MiniMax-M3 is released with open weights, so it can be self-hosted.
Compare MiniMax-M3
Methodology: Score is 0 to 100 and computed, not typed: 55% independent evals (Artificial Analysis, Arena, Scale SEAL, Epoch, each scaled against its natural floor and the best score in this table, then averaged), 15% blended price on a fixed log scale ($0.05 per million is 100, $60 is 0), 15% practitioner sentiment from Signals and community threads, 15% operational fit (context window, open weights). A missing component drops out and the rest are reweighted. A model with no independent eval yet is provisional and ranks below every measured one. Within three points is a tie. Vendor-published figures are labelled and never counted.