MiniMax Group Inc. has introduced the M3.1 Flash model, an update to its M3 large language model. The new version is designed to reduce the costs of running AI agents by 70% and offers inference speeds nearly three times faster than its predecessor, reaching peak output of 200 tokens per second.
The model features 428 billion parameters and utilizes aggressive sparsity at inference time to improve performance. It supports text, image, and video inputs with a context window of up to 1 million tokens. Developers can adjust reasoning effort across five levels, ranging from Low to Max.
In internal benchmarks, M3.1 Flash achieved a score of 61.7 on OSWorld V2 and 76.2 on BankerToolBench. For security-related tasks, such as code auditing and malware analysis, the model scored 82% on the CyberGym benchmark. The company also confirmed that the previously anonymous model Space Bunny was an early build of M3.1 Flash.
Pricing for the model is set at $0.15 per million input tokens, with output costs at $0.60 per million tokens and cached input at $0.015. The model is currently accessible via the MiniMax API, the M Plan subscription, and the MiniMax Code desktop application, with a more advanced Pro version expected in the future.
