MiniMax Debuts M3.1 Flash to Slash AI Agent Operating Costs

MiniMax Group Inc., a Chinese artificial intelligence developer, unveiled M3.1 Flash on Tuesday, positioning the new model as a leaner and quicker successor to the M3 large language model that debuted in June. The offering directly addresses the financial burden of deploying AI agents, which typically require numerous sequential model invocations to execute complex workflows. According to MiniMax, M3.1 Flash embodies its "ongoing exploration of the balance between efficiency and capability." The company claims that agentic workloads now cost 70% less to operate compared to its predecessor, while inference speed has improved nearly threefold, reaching peak output rates of 200 tokens per second during peak usage windows.

The parameter count remains largely consistent with M3, holding steady at 428 billion parameters according to MiniMax's developer documentation for the preview build. The performance gains stem from heightened sparsity applied during inference operations. Reasoning capabilities remain active across all requests, with developers able to select from five intensity settings ranging from Low through Max. The model processes multimodal inputs including text, images and video content, and supports context windows extending to 1 million tokens.

In internal benchmarking, M3.1 Flash achieved a score of 61.7 on OSWorld V2, a test measuring how effectively models can interact with computer interfaces. MiniMax asserts this represents the highest performance among Chinese-developed models on this metric. On BankerToolBench, which evaluates banking workflow capabilities, the model scored 76.2. For cybersecurity applications, M3.1 Flash registered an 82% result on CyberGym, a benchmark developed by University of California, Berkeley researchers to assess whether AI agents can identify and reproduce genuine software vulnerabilities. MiniMax highlighted code review, penetration testing and malware detection as among the security-focused tasks the model can handle.

Pricing for M3.1 Flash stands at 15 cents per million input tokens, matching the off-peak rate charged by Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co. Ltd. for its V4.1-Flash variant. DeepSeek reported that its updated V4.1-Flash resolved 77% of tasks on the DeepSWE v1.1 software engineering benchmark, while M3.1 Flash achieved 73.7% on the same test. Neither result has undergone independent verification. MiniMax's share price declined following DeepSeek's initial V4.1-Flash release in September, according to Bloomberg reporting.

The announcement also clarifies the identity behind Space Bunny, an unattributed model that surfaced on OpenRouter Inc.'s model routing platform in late September. Community members identified the model as MiniMax's work by analyzing its tokenizer characteristics, and the company confirmed today that Space Bunny represented an earlier iteration of M3.1 Flash. The anonymous version processed more than 200 trillion tokens before its creator was revealed, per MiniMax's launch statement.

M3.1 Flash is now accessible via the MiniMax API, the company's M Plan subscription tier and its MiniMax Code desktop application. Output pricing is set at 60 cents per million tokens, with cached input charged at 1.5 cents per million tokens. MiniMax indicated that model weights will become available at a later date, and signaled that a higher-capacity Pro variant is under development. MiniMax has been publicly traded on the Hong Kong exchange since completing a $619 million initial public offering in January.