China's AI company MiniMax has launched M3.1 Flash, a large language model (LLM) that increases inference speed and sharply reduces usage costs. It is a strategy to compete with global companies in the AI agent market by leveraging price competitiveness.

On Oct. 19, local time, IT outlet SiliconANGLE reported that M3.1 Flash is a follow-up to M3, which launched in June. MiniMax said it cut AI agent task costs by 70 percent compared with the previous model and raised inference speed by about threefold.

The model adopts a mixture-of-experts (MoE) architecture with a total of 428 billion parameters. It improved efficiency by strengthening sparsity technology that activates only some parameters during inference.

The company said output speed is more than 200 tokens per second, and exceeds 170 tokens on a median basis.

It supports text, image and video inputs and can process up to 1 million tokens of context. Inference is enabled by default, and developers can adjust inference intensity across five levels.

The company also highlighted competitiveness in performance evaluations. It scored 61.7 on OSWorld V2, which measures computer operation capability, and 76.2 on BankerToolBench, a financial task evaluation. It scored 82 percent on the cybersecurity evaluation CyberGym.

It scored 73.7 percent on DeepSWE v1.1, which assesses software development capabilities. That is lower than the 77 percent disclosed for rival DeepSeek's V4.1 Flash, but both figures were released by the developers and were not independently verified.

Pricing is $0.15 per 1 million input tokens and $0.6 per 1 million output tokens. Cached input is priced at $0.015.

MiniMax said earlier that Space Bunny, which was run anonymously on OpenRouter, was an early version of M3.1 Flash. It said the number of tokens processed during the anonymous operation exceeded 200 trillion.

It is currently available through the MiniMax API, subscription service and a coding app. The company plans to disclose model weights in the future and also introduce a high-performance Pro version.