MiniMax-M3.1-Flash: Cheaper and Smarter, but It Hallucinates
- Source
- x.com
- Date
MiniMax-M3.1-Flash scores 42 on the Artificial Analysis Intelligence Index, 13 points higher than MiniMax-M3 and less than half the cost per task. It scores level with GLM-5.3-Flash but behind DeepSeek V4.1 Flash 1017 MiniMax-M3.1-Flash is @MiniMax_AI's new Flash reasoning model, which we benchmarked at its max effort setting. It scores 42 on the Artificial Analysis Intelligence Index, level with GLM-5.3-Flash (max, 42), just ahead of Gemini 3.8 Flash (high, 41), and behind DeepSeek V4.1 Flash 1017 (max, 45) Key takeaways: ➤ Complex reasoning improved substantially compared to M3: MiniMax-M3.1-Flash scores 48% on Humanity's Last Exam, ahead of GLM-5.3-Flash (max, 40%) and DeepSeek V4.1 Flash 1017 (max, 43%), and level with Gemini 3.8 Flash (high, 48%). On CritPt it scores 22%, ahead of GLM-5.3-Flash (15%) and Gemini 3.8 Flash (18%) and just behind DeepSeek V4.1 Flash 1017 (24%). These are up from MiniMax-M3 which scored 39% and 4% respectively ➤ Fewer output tokens, lower cost per task: MiniMax-M3.1-Flash used ~189M output tokens to run the Intelligence Index, considerably fewer than GLM-5.3-Flash (max, ~324M), DeepSeek V4.1 Flash 1017 (max, ~224M) and Gemini 3.8 Flash (high,…
Context
Artificial Analysis reports that MiniMax's new MiniMax-M3.1-Flash, a it tested at the maximum effort setting, scores 42 on its Intelligence Index, a composite of the firm's evaluations (Artificial Analysis on X). That is 13 points above MiniMax-M3, level with GLM-5.3-Flash, one point ahead of Gemini 3.8 Flash on high effort, and behind DeepSeek V4.1 Flash 1017 at 45. The clearest improvement is on harder reasoning tests: the model scores 48% on Humanity's Last Exam and 22% on CritPt, a test Artificial Analysis groups under scientific reasoning, up from 39% and 4% for MiniMax-M3.
The cost story is about how much the model writes, not only its per-token price. Artificial Analysis counts roughly 189M output tokens for MiniMax-M3.1-Flash to run the full index, compared with about 224M for DeepSeek V4.1 Flash 1017, 324M for GLM-5.3-Flash and 334M for Gemini 3.8 Flash. Combined with list pricing of $0.15 per million input tokens and $0.60 per million output, that works out to about $0.22 per index task, against roughly $0.26 for the DeepSeek model, $0.28 for GLM-5.3-Flash, $1.24 for Gemini 3.8 Flash and $0.51 for MiniMax-M3.
The weak spot is AA-Omniscience, a factual-recall test that, under Artificial Analysis's methodology, rewards correct answers, subtracts for wrong ones and does not penalize a model for saying it does not know. MiniMax-M3.1-Flash gets 56% of questions right, well above GLM-5.3-Flash at 28%, but it attempts almost every question. Its rate, which measures how often a model gives a wrong answer when it should have declined, is 93%, so nearly everything it does not know comes out as a confident error, holding its overall Omniscience score to 16. The same-day DeepSeek write-up shows the opposite trade: that update's Omniscience gain came largely from declining to answer more often.
The model takes text and image input, has a 1M-token and is available through MiniMax's own API. Artificial Analysis describes it as closed weights for now, with MiniMax announcing an open-weights release on October 23.
- reasoning model — A model trained to think — generating extended internal reasoning before answering — trading time and tokens for accuracy on hard problems.
- hallucination — When a model states something false with full confidence — inventing facts, citations, or APIs that don't exist.
- context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
- open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.
Checking sign-in…
Loading comments…