Minimax M3
MinimaxFlagship
ThinkingTool UseVisionStructured Output
About this model
MiniMax-M3 is a multimodal foundation model from MiniMax. It supports text, image, and video inputs with text output, a 1M-token context window, and is suited for long-horizon agentic work, coding,...
Performance Tier
Flagship
Minimax M3 is a flagship model from Minimax : the most capable in their lineup.
Best-in-class model from this provider. Highest performance across benchmarks, ideal for demanding tasks.
Pricing
This model is included in Elosia plansAffordable
Low cost. Suitable for sustained use and high-volume interactions.
| Type | per 1M tokens |
|---|---|
| Input (prompt) | $0.300 |
| Output (completion) | $1.20 |
| Cache read | $0.060 |
Capabilities
Context Length524K
Max Output Tokens512K
TokenizerOther
Inputtext, image, video
Outputtext
Release DateMay 31, 2026
Benchmarks
General Intelligence
MMLU
Not reported
GPQA Diamond
Not reported
Mathematics
MATH-500
Not reported
Programming
HumanEval
Not reported
SWE-bench Verified
Not reported
Reasoning
IFEval
Not reported
Agentic
SWE-bench Pro
59%
Recommended Use Cases
CodingAnalysisResearchData Extraction
Strengths
- MoE architecture (~428B total, 23B active) — strong open-weight agentic coding (SWE-bench Pro 59.0), ahead of GPT-5.5 and Gemini 3.1 Pro
- 1M-token context made efficient by MiniMax Sparse Attention (MSA) — ~9× prefill and ~15× decode speedup over the prior generation at 1M tokens
- Native multimodal input including video — text, image and video from the first training step, plus computer-use workflows
- Cost-efficient at $0.30/M input and $1.20/M output — a fraction of comparable frontier models for agentic coding
- Open-weight with a full technical report (arXiv 2606.13392) — downloadable and self-hostable
Limitations
- Headline benchmarks are self-reported by MiniMax; independent verification remains limited
- SWE-bench Pro 59.0 trails the leading closed model (Claude Opus 4.8, ~69) on pure code modification
- ~428B MoE with a custom MSA operator makes self-hosting heavyweight; contexts above 512K bill at a higher tier
Resources
This model may use your data for training
