🔍 Read the full analysis: Does The Open-Weight Leaderboard Tell The Whole Story About MiniMax M3.1 Flash? on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
MiniMax released M3.1 Flash on October 19, 2026, at $0.15 per million input tokens and $0.60 per million output tokens, while saying its weights will arrive October 23. Artificial Analysis gives it an Intelligence Index score of 43, behind several open-weight competitors, and measures lower speed than MiniMax claims. The supplied source describes a 24-article newsroom test but cuts off before reporting its scores, so those results cannot be assessed here.
MiniMax released M3.1 Flash on October 19, pricing it at $0.15 per million input tokens and $0.60 per million output tokens and pitching it as a fast, low-cost model for agent tasks. An independent Artificial Analysis index scores it at 43, below several open-weight models, while recording slower output than MiniMax’s advertised speed; the weights are not yet public, so the model is currently listed as proprietary.
MiniMax says M3.1 Flash costs half as much per token as its predecessor, M3, and about 70% less per agent task. Those are company claims, not results established by the independent index. The company also advertises everyday output above 200 tokens per second and a median above 170, and says the model is nearly three times faster than M3. Its published design description is limited to “more aggressive sparsity.”
Artificial Analysis measures 120 output tokens per second and a 1.08-second time to first token. Its Intelligence Index score of 43 trails Xiaomi’s MiMo-V2.6-Pro at 46, DeepSeek V4.1 Flash 1017 and GLM-5.3 at 45, and Step 5 Preview and Kimi K3 at 44. The index is a standardized comparison, but it does not by itself determine how well a model will perform on every real-world task.
The source report also describes a newsroom evaluation using 24 articles in English and German, drawn from 12 news sources and scored by two independent model-based judges. It explains the test design but provides no score table or findings in the supplied material. No conclusion about M3.1 Flash’s newsroom accuracy or writing quality can be drawn from that test here.
Infographic 01 · MiniMax M3.1 Flash
A step behind on the index. Ahead of every open model in our newsroom.
An independent intelligence index puts MiniMax M3.1 Flash a step behind the open-weight leaders. Writing real news from real sources, it beat all of them.
Independent index
Artificial Analysis Intelligence Index, 10 evaluations. Bar length on a 0–50 scale.
Artificial Analysis lists M3.1 Flash as proprietary until its weights are released.
Our newsroom test
Mean score, English articles, 1–5 by GPT-6.1 Sol and Claude Opus 5.5 with the source in context. Bar length on a 0–5 scale.
In German, M3.1 Flash (3.36) also leads every open model, but GPT-6 Luna is clearly ahead (3.62).
Sources: Artificial Analysis, MiniMax-M3.1-Flash and open-weights models (20 Oct 2026); Thorsten Meyer AI newsroom benchmark, 12 sources × English + German, run 20 Oct 2026.
Powered by Thorsten Meyer AI
Price Versus Measured Performance
The launch puts a low listed token price beside a more mixed performance picture. For developers choosing a model for agent workflows, the relevant cost is not only the price per token but also how many tokens the model uses and how quickly it completes a task. Artificial Analysis says M3.1 Flash used 180 million tokens to complete its index evaluation, compared with a 99 million-token median across models. That figure is specific to the index run, not a direct measure of ordinary customer costs, but it cautions against treating a low per-token rate as a complete cost comparison.
The difference between company speed claims and index measurements also matters for applications where response time affects users. The two figures may reflect different settings or measurement methods; the available material does not explain the gap. Until more details are published, readers should treat the company’s headline speed and the index’s measured speed as distinct results rather than interchangeable estimates.
For newsrooms and other organizations that rely on faithful summarization or drafting, leaderboard scores are only part of the decision. A model can rank well on general evaluations yet still make errors on particular source-based tasks. The reported newsroom test could address that question, but its missing results prevent readers from judging how M3.1 Flash handled factual accuracy, invented details, clarity, or structure.
AI language model performance evaluation tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
From Space Bunny to M3.1
Before its public naming, the model reportedly ran on OpenRouter as Space Bunny Alpha. The source says it appeared anonymously on September 23, led OpenRouter’s usage ranking for 10 days and processed 56.7 trillion tokens there before being removed on October 5. MiniMax separately counts more than 200 trillion tokens overall. These platform and company figures refer to different scopes and should not be combined as though they measured the same usage.
MiniMax revealed the model as an early version of M3.1 Flash on October 19 and made it available through its API, M Plan subscription, MiniMax Code and OpenRouter. OpenRouter offered free access for the first 72 hours, according to the source. MiniMax says the weights will be released on October 23; until that happens, developers cannot inspect or run the model’s weights as an open-weight release.
MiniMax’s own benchmark charts compare M3.1 Flash with selected competitors. In those results, it scores 62.2 on Automation-Bench, ahead of DeepSeek V4.1 Flash 1017 at 60.5, but DeepSeek leads on DeepSWE and Terminal-Bench. The figures are self-reported comparisons; MiniMax selected the rivals, and some newer competing models are absent. The source also notes that MiniMax’s “first among Chinese models” claim for Agents’ Last Exam omits DeepSeek from that chart row, while DeepSeek’s model card reports a different score that may use another setup.
““more aggressive sparsity””
— MiniMax
As an affiliate, we earn on qualifying purchases.
Weights, Data and Test Results
The weights have not been released, and the available source gives neither an official license nor a confirmed parameter count. A reported figure of 428 billion total parameters and 23 billion active parameters is not attributed to MiniMax and is explicitly unconfirmed; it should not be treated as established fact.
There is also an unresolved data question from the model’s anonymous OpenRouter period. The source says the provider could retain prompts but could not use them to train the model. MiniMax has not said what happened to that data. The source does not establish whether prompts were retained, how long they may have been kept, or what policies apply now.
MiniMax’s benchmark and security figures, including its reported CyberGym result, come from company evaluations and have not been independently reproduced in the supplied material. Artificial Analysis provides an external comparison, but differences in benchmark setup can affect scores. Finally, the newsroom test’s actual results are absent from the source excerpt, leaving its central performance findings unknown.
As an affiliate, we earn on qualifying purchases.
The October 23 Weight Release
The next stated milestone is MiniMax’s planned October 23 release of the weights. The release, if it occurs as promised, should clarify the model’s license and permit outside users to inspect and test the weights. The source does not confirm a parameter count or describe what documentation will accompany the release.
Further independent evaluations could help explain the differences between MiniMax’s advertised speed and Artificial Analysis’s measurement, and test whether results hold across other tasks and configurations. Readers should also look for the missing newsroom benchmark results before drawing conclusions about the model’s reliability in source-based reporting. MiniMax has not supplied an answer in the provided material about prompts from the anonymous OpenRouter period.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is MiniMax M3.1 Flash?
M3.1 Flash is a model MiniMax released on October 19, 2026. The company describes it as a low-cost, fast option for agent tasks and offers it through several services, including its API and OpenRouter.
Is M3.1 Flash open-weight now?
No. The weights were not public at the time covered by the source. MiniMax said they would arrive on October 23, but the source does not confirm that release or provide a license.
How does it rank against open-weight competitors?
Artificial Analysis gives M3.1 Flash an Intelligence Index score of 43. Several listed competitors score higher, including MiMo-V2.6-Pro at 46 and DeepSeek V4.1 Flash 1017 and GLM-5.3 at 45.
Is it as fast as MiniMax claims?
MiniMax claims everyday speeds above 200 tokens per second. Artificial Analysis measured 120 tokens per second in its evaluation. The source does not explain the difference in measurement conditions, so the figures should be kept separate.
What did the newsroom test find?
The supplied material describes a test of 24 English- and German-language articles, but it ends before giving the results. It therefore does not establish how well M3.1 Flash performed on factual fidelity, invented details, or writing quality.
Source: ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
