Overview
- Xiyi Tech has rolled MiniMax M3.1 Flash into its product stack, offering access via MiniMax API, MPlan and MiniMax Code while saying the full model weights will be published later.
- The company reports near 3× faster online inference and about a 70% drop in per-task cost compared with its prior generation.
- Xiyi published per‑million‑token prices of input 1 yuan, output 4 yuan and cache read 0.1 yuan to show the model’s commercial billing structure.
- Company-provided benchmark results show leading scores on tests for agentic tasks, coding, computer use and security (including BankerToolBench, DeepSWE, OSWorld, ALE and CyberGym) but these results have not been independently verified.
- Xiyi says an early anonymous deployment called Space Bunny processed over 200 trillion tokens as in-field evidence, a claim that supports its rollout but that customers should validate through weight release and independent testing before wide adoption.