MiniMax M3.1 Flash targets multi-step work: 72.5% AutomationBench-AA, best open model
vllm_project · x · 2026-10-19
Part 2 of vLLM's Day-0 series highlights MiniMax M3.1 Flash's agentic and coding results:
- 72.5% on AutomationBench-AA — leading open models in its launch comparison, covering co-work automation
- 73.67% on DeepSWE — real-codebase understanding, bug fixes and feature development
The model is positioned for workflows that require planning, tool use and carrying tasks through, rather than single-turn responses.
More from coding & agent
- vLLM 0.32.0 adds structured-decision API reading first-token probabilities — AccBalanced · 2026-10-19
- 10 open-source GitHub repos for fully local speech-to-text, led by Whisper at 110K stars — victor_explore · 2026-10-19
- All Things Open talk: hardening agent security with Docker sandboxes — HowDevelop · 2026-10-19
- Nace AI launches Drex Max, a screenshot-in decision model for agents at 174 ms and $0.06/M tokens — testingcatalog · 2026-10-19
- Watch Claude Opus 5.5 build a game-ready bush via Blender MCP in 30 seconds — sidahuj · 2026-10-19
- Anthropic: Engineers' New Role Is Saying "No" to Claude — But Who Trains the Next Generation? — sqcai · 2026-10-19