Overview
- Xiyu says it has deployed MiniMax M3.1 Flash into its MiniMax API, MPlan and MiniMax Code toolchain and will release model weights at a later date for outside validation.
- The company reports near 3x faster online inference and about a 70% reduction in per-task cost, and it published per‑million‑token pricing of ¥1 for input, ¥4 for output and ¥0.1 for cache reads.
- Xiyu provided benchmark results the firm says show strong capability: BankerToolBench 76.2, DeepSWE v1.1 73.7, OSWorld V2 61.7, ALE 29.6 and CyberGym 82%, which it links to improvements in coding, long‑chain task execution and computer‑use skills.
- The firm says an early anonymous deployment called 'Space Bunny' processed more than 200 trillion tokens and that M3.1 Flash can run multi‑step enterprise workflows across sales, operations, finance and HR and support security tasks such as code audit, vulnerability discovery and penetration testing.
- Because the benchmarks and usage claims come from Xiyu's internal tests and the weights are not yet public, independent verification is still required and wider market uptake will depend on third‑party reproduction and real customer deployments.