MiniMax M3.1 Flash Official Evaluation ReportProvider report
2026-10-20All 15 MiniMax M3.1 Flash results from the official release table; comparison models excluded. Automation-Bench uses the v1.0.6 public 600-task set. ALE uses the official Overall full set, and OSWorld 2.0 uses 0808 Full; CLI-only and offline subsets are kept separate. NL2Repo uses corrected tests and is recorded separately. HLE uses 2,158 text-only questions without tools. CyberGym uses the table's precise 82.2 rather than the prose's 82%. GDP.pdf and ALE quote third-party results. Per-metric harness versions, context settings and source discrepancies are preserved in raw records.