MiniMax launches first Flash model M3.1-Flash-Preview with 1M context and five reasoning tiers
赛博禅心 · wechat · 2026-10-01
MiniMax launched M3.1-Flash-Preview on Sept 27, its first Flash model, available in MiniMaxCode and via API for TokenPlan subscribers.
Specs: 1M token context, native multimodal input (text/image/video), deep thinking in five tiers (low to max, default max). Still a preview with no published benchmarks. An anonymous model "SpaceBunnyAlpha" on OpenRouter since Sept 23 is widely speculated to be its gray-release version, with inferred changes including a 524K completion cap and tuning toward fast coding and long-running agents.
Highlights from early testing:
- Fixed real Luxon date-library issues via PRs (including DST and float-precision edge cases) and a SQLite multi-process write-lock bug in MiniMaxCode's repo
- Perf optimization: a slow ticketing system went from 15.3s to 0.045s for list load
- Frontend/3D: SVG pelican riding a bike with IK-driven legs, Three.js procedural Temple of Heaven, a 17,522-voxel Millennium Falcon, a working Windows XP replica
- Video understanding: split a 1m51s clip into a 58-shot storyboard with frame-accurate timecodes
- Finance: one prompt produced a full 3-statement DCF model for Kweichow Moutai
More from coding & agent
- Dev Has Claude Build a Custom Jump Animation Editor to Speed Up Game Feel Iteration — andrew_n_carr · 2026-10-02
- Translating an entire book with DeepSeek: pennies and under an hour, decent quality — teortaxesTex · 2026-10-02
- OmniSeek turns Omni-LLMs into agents that actively seek audio-visual evidence — Haibo Wang · 2026-10-02
- Microsoft's ActiveSaddler Uses Automated Curriculum Learning to Boost Agent Harnesses by 7.5 Points — microsoft · 2026-10-02
- Alibaba's PoS Maintains Explicit Belief States to Fix Long-Horizon Agent 'Belief Trapping' — alibabagroup · 2026-10-02
- IntentFlux Benchmarks 'Intent Drift' in LLM Agents: Scores Fall from 0.476 to 0.384 as Users Change Their Minds — Yanjie Zhang · 2026-10-02