MiniMax-M3.1-Flash
Released Oct 19, 2026 · knowledge cutoff unpublished
- Status
- Unknown
- Location
- China
- Type
- LLM
- Modality
- Multimodal
- Context window
- 1M
- Max output
- —
- Speed
- —
- Price ($/MTok in / out)
- $0.15 / $0.60
- Cost per task
- $0.22
- Open weights
- No
Strengths & weaknesses
Strengths
- Already battle-tested at scale before its name was public — as the anonymous 'Space Bunny' it was OpenRouter's most-used model for ten straight days, so the launch build is not a cold start
- Built around agent work rather than chat: MiniMax pitches computer-use and cross-application tasks first, and its Code Arena WebDev debut jumped well clear of MiniMax-M3
- Reasoning model priced like a Flash — cheap enough to leave running inside long agent loops, with MiniMax claiming roughly three times faster online inference than M3
Weaknesses
- Very verbose: Artificial Analysis' index run drew nearly twice the median output tokens, so the low per-token rate buys less than it looks
- Almost never declines to answer — on Artificial Analysis' knowledge test it attempted nearly every question it did not know, so unfamiliar facts come back as confident guesses
- Weights are promised but not yet out, and MiniMax published no parameter count or tracked-benchmark scores at launch — only its own agentic suites
Announced 2026-10-19 on MiniMax's blog and X account as the public launch of M3.1 Flash, after a MiniMax Code / Token Plan-only preview (M3.1-Flash-Preview, 2026-09-27) and the 'Space Bunny' OpenRouter stealth run, which MiniMax says was an earlier version of this model; available the same day through the MiniMax API (pay-as-you-go), M Plan, MiniMax Code and OpenRouter. MiniMax reports only its own agentic suites — OSWorld V2 61.7, Agents' Last Exam 29.6, DeepSWE v1.1 73.7%, BankerToolBench 76.2 and CyberGym ~82% — none of the site's tracked benchmarks, and says HLE edges DeepSeek V4.1 Flash without giving a figure, so every benchmark cell is null; Arena's Code Arena WebDev listing (1588) is not the text-arena Elo this site records. Artificial Analysis rates it 42 on its Intelligence Index (MiniMax-M3: 29) at $0.22 per index task, measured on the single unlabelled variant it lists, hence effort null; speed stays null because AA's output-speed figure could not be confirmed. Pricing is $0.15 input / $0.015 cache read / $0.60 output per million tokens, with a priority tier at 1.5x. Context is 1M tokens shared by input and output; MiniMax's Messages API reference gives a 524,288-token output maximum but names the Preview model id, so maxOutput stays null. Parameter count is unpublished. Weights are promised for 2026-10-23 per MiniMax's X post (the blog says only 'later'), so openWeights stays false and license null until they ship. predecessorId is null: MiniMax names no model it replaces.
Benchmarks0 of 10 reported
MMLU-Proretired
GPQA Diamond
SWE-bench Verified
SWE-bench Pro
Terminal-Bench 2.1
AIME (latest)retired
Humanity's Last Exam
LMArena Elo
GDPval-AA v2
ARC-AGI-2
For developers
API model strings
- minimax
MiniMax-M3.1-Flash
Licence
Proprietary — weights not released
Retirement
No retirement announced
Lineage
No recorded predecessor