MiniMax launches M3.1 Flash model, says it cuts AI agent costs by 70%

Chinese AI developer MiniMax Group Inc. has unveiled M3.1 Flash, a more efficient and affordable update to its M3 large language model, announced on June. MiniMax claims that this new model reduces the cost of running AI agents by 70% compared to its previous model, M3. The inference speed of M3.1 Flash is nearly three times faster, with an output of over 200 tokens per second at peak performance.

The model maintains 428 billion parameters, similar to its predecessor, with enhanced sparsity at inference time being the primary reason for its increased speed. Developers can choose from five effort levels, ranging from Low to Max, to adjust the reasoning process for each request. M3.1 Flash supports text, image, and video inputs, with prompts capable of reaching up to 1 million tokens.

In MiniMax's tests, it scored 61.7 on OSWorld V2, a benchmark for computer operation tasks, outperforming other Chinese models. The model also achieved 76.2 on BankerToolBench, a banking workflow test, and 82% on CyberGym, a security test. For the University of California, Berkeley's benchmark, M3.1 Flash delivered 73.7% on code auditing, penetration testing, and malware analysis jobs.

Pricing for M3.1 Flash is set at 15 cents per million input tokens, similar to Hangzhou DeepSeek's V4.1-Flash model. MiniMax's shares dropped after the initial release of V4.1-Flash in September, according to Bloomberg. The launch of M3.1 Flash also reveals that Space Bunny, an anonymous model found on OpenRouter Inc.'s model routing service, is an early version of MiniMax's M3.1 Flash.

M3.1 Flash is currently available through MiniMax's API, M Plan subscription, and MiniMax Code desktop app. Output costs 60 cents per million tokens, while cached input is priced at 1.5 cents. MiniMax plans to release the Flash weights later and is also developing a more advanced Pro model.