$ cat wiki/models/minimax-m3-1-flash.md
MiniMax M3.1 Flash
Spec
| Attribute | Value |
|---|---|
| Developer | MiniMax |
| Released | 2026-10-19 |
| Announced | 2026-10-19 |
| Context window | unknown |
| Pricing | $0.15/M input · $0.60/M output · $0.015/M cached input |
| License | unknown |
| Availability | MiniMax API, M Plan, MiniMax Code; weights promised later |
| The launch supplies these rates and access channels but does not state a context window or license. Weights are a future release, not an available download established here. (source) |
Release Date
Released 2026-10-19. MiniMax identifies the anonymously served Space Bunny as an early version; the announcement does not give a date for releasing weights. (source)
Benchmarks
Vendor-reported results: DeepSWE v1.1 73.7%, BankerToolBench 76.2, OSWorld V2 61.7 points, Agents’ Last Exam (ALE) 29.6 points, and CyberGym 82%. The prose supplies no reasoning-effort or harness settings for these scores. Graphical comparisons and independent evaluations were not reviewed. (source)
MiniMax reports output speeds reaching 200+ TPS, with P50 170+ TPS, and 70% lower overall agent-task usage cost than its predecessor. These are vendor workload claims, distinct from the published token rates. (source)
Use Cases
MiniMax positions the model for multi-step enterprise workflows, software engineering, computer use and security work. Those are vendor use cases, not evidence that arbitrary production tasks are reliable. See Agents (LLM Agents). (source)
Compared To
The earlier MiniMax M3 remains a separate model record. This launch adds an explicitly priced Flash deployment; its claim of lower task cost does not specify enough workload detail to reproduce that comparison. See Eval Harness Configuration. (source)
Sources
- MiniMax announcement, captured in the snapshot.