What was announced
MiniMax has made its new multimodal model M3.1 Flash available through its API. The model uses a mixture-of-experts (MoE) architecture with sparse attention and a native visual encoder. Of its 428 billion total parameters, only about 23 billion activate per token.
Key features
- 1 million token context window
- Text, image and video input; text output
- Always-on reasoning with 5 effort levels (Low, Medium, High, Xhigh, Max), Max by default
- Pricing: $0.15 per 1M input tokens, $0.60 per 1M output tokens
Performance
According to MiniMax's own table, the model stands out on agentic office work, scoring 62.2 on Automation-Bench v1.0.6, ahead of Opus 5 (50.3) and GPT-5.6 Sol (45.8). Coding results are mixed: it scores 73.7 on DeepSWE v1.1 but only 38.1 on Terminal Bench 4.0. All scores are company-reported with no independent replication yet.
Notes for developers
Reasoning cannot be disabled; requests sending thinking.type=disabled or effort=none return an error. Responses arrive in a separate reasoning_content field. MiniMax confirmed that the stealth model Space Bunny was an earlier M3.1 Flash build and processed over 200 trillion tokens. Model weights are expected on October 23, 2026.
