MiniMax M3
M3 is a native multimodal model trained with mixed modalities from the start. It uses MiniMax Sparse Attention for long context and accepts text, image and video input.
One release with several deployment formats
The official local guide uses an eight-B200 reference configuration with MXFP8 weights. That is a documented deployment path, not a universal minimum. The MXFP8 download is kept with M3 rather than presented as a separate trained model.MiniMax local deployment overview
Choose the reasoning behavior you need
Practical guidance. M3 is worth a separate comparison when an application needs a direct response mode. Measure that mode on the same requests as adaptive reasoning, and check whether reduced delay changes answer quality.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Launch date | |
| Access | API, Downloadable, Provider product |
| Model identifiers | MiniMax-M3MiniMax-M3-MXFP8 |
| Context window | 1,000,000 tokens Native context capacity published for M3. MiniMax M3 model card |
| Architecture | About 428 billion total parameters and 23 billion active parameters.MiniMax M3 model card |
| Thinking modes | Enabled, adaptive or disabled. This differs from M3.1 Flash, whose API requires thinking.MiniMax M3 model cardMiniMax text model invocation and thinking controls |
| Weight terms | The Community License requires attribution and a commercial-use notice. Above US$20 million yearly product or service revenue, prior written authorization is required.MiniMax M3 Community License |
Related models
- MiniMax M3.1 Flash
MiniMax · MiniMax M3
A multimodal coding and agent model with sparse attention, quantization-aware training and a DSpark draft head.
Launch date API · Downloadable · Provider product - MiniMax M1 80K
MiniMax · MiniMax M1
A Text-01-derived reasoning model trained with an 80,000-token thinking budget and tool-call support.
Launch date not confirmedDownloadable