MiniMax M3.1 Flash: What We Learned From Testing It Against Open Models

🔍 Read the full analysis: MiniMax M3.1 Flash: What We Learned From Testing It Against Open Models on ThorstenMeyerAI.com

MiniMax released M3.1 Flash on October 19, pricing the model at $0.15 per million input tokens and $0.60 per million output tokens. A review by ThorstenMeyerAI.com found a mixed performance picture: Artificial Analysis scored it below several open-weight competitors, while the review also tested it on 24 real news articles, but the supplied report excerpt does not include the newsroom test’s results.

MiniMax describes M3.1 Flash as a low-cost, fast model for agent work. The company says it is nearly three times faster than its M3 predecessor and can exceed 200 tokens per second in everyday use. Artificial Analysis measured 120 tokens per second and a 1.08-second time to first token in its own evaluation. Those figures come from different sources and testing methods, so they are not directly interchangeable.

The model has a one-million-token context window, adjustable reasoning effort and a cached-input price of $0.015 per million tokens. MiniMax says the model uses “more aggressive sparsity,” but has not published an official parameter count. Its weights were not available at the time of the report; MiniMax said they would arrive on October 23. A reported parameter count from another outlet remains unconfirmed by the company.

Artificial Analysis gave M3.1 Flash an Intelligence Index score of 43. In the comparison cited by the review, DeepSeek V4.1 Flash 1017 scored 45, as did GLM-5.3, while MiMo-V2.6-Pro scored 46. The review also ran 24 English- and German-language articles based on 12 news sources, with two independent judges assessing fidelity, fabricated details, clarity, usefulness, structure and polish. The available source material describes the methodology but does not show the resulting scores or comparisons.

Price Versus Tested Performance

M3.1 Flash’s headline appeal is its low per-token price, set at half the listed input and output rates of MiniMax M3. MiniMax also claims the new model costs about 70% less per agent task, but that is a company estimate rather than an independently verified result. For businesses running agents at scale, the difference between token price and task cost matters: a model that uses more tokens can narrow the savings.

Artificial Analysis reported that M3.1 Flash used 180 million tokens to complete its index evaluation, compared with a median of 99 million. That is a raw comparison from the index’s test, not a measure of cost across every workflow. Still, it points to a practical question for buyers: whether lower rates per token translate into lower total spending for their own prompts and tasks. The review’s missing newsroom scores also mean readers cannot assess how the model performed against alternatives on that test.

The results also complicate MiniMax’s speed pitch. The company’s claim of more than 200 tokens per second is higher than Artificial Analysis’s 120-token-per-second measurement. Differences in conditions may account for some of the gap, but the source material does not establish why the figures diverge. Customers comparing models should treat both as attributed measurements rather than a single settled speed figure.

From Space Bunny to M3.1

Before MiniMax identified the model, it appeared anonymously on OpenRouter as Space Bunny Alpha. According to the review, it went live on September 23, ranked as OpenRouter’s most-used model for 10 days and processed 56.7 trillion tokens on the platform before being removed on October 5. MiniMax says the model processed more than 200 trillion tokens overall. These usage figures are attributed to the review and company, respectively.

MiniMax said prompts could be retained during the anonymous period, though not used for training. The company has not explained what happened to that data. The episode may matter to users who tried the model before its identity was disclosed, particularly organizations whose prompts contained sensitive information.

MiniMax’s own benchmark chart, transcribed by CellCog for the review, showed M3.1 Flash ahead of DeepSeek V4.1 Flash 1017 on Automation-Bench, 62.2 to 60.5. DeepSeek scored higher on DeepSWE and Terminal-Bench. The review cautioned that MiniMax selected the competitors and that the results were self-reported. It also noted that MiniMax’s “first among Chinese models” claim on Agents’ Last Exam omitted DeepSeek from the chart row, while DeepSeek’s model card reports a different score, potentially from a different setup.

„“nearly three times faster” than M3“

— MiniMax

Missing Scores and Data Details

The supplied report excerpt does not give the English and German scores from the 24-article newsroom test, nor a completed ranking against the other models. Its methodology is described, but without the results it is not possible to say from this material whether M3.1 Flash performed well, poorly or differently by language.

Other open questions include the model’s official parameter count, the license and the eventual terms for accessing its weights. MiniMax’s promised October 23 release had not yet occurred at the time described. The anonymous OpenRouter period also leaves a data-handling question: MiniMax has not said what happened to retained prompts. Benchmark comparisons remain limited by differing test setups, and the company’s own chart results have not been independently reproduced in the source material.

Weights and Test Results Ahead

The next stated milestone is MiniMax’s promised October 23 release of the weights. Whether that release happens as scheduled, what license applies and whether the model’s architecture details are published remain unconfirmed in the report. Weight availability would also change how the model is classified by independent trackers that listed it as proprietary while the weights were unavailable.

For a fuller assessment, readers will need the newsroom benchmark’s missing language-by-language results and comparable results from other models tested under the same prompt and judging process. Further independent measurements could also clarify the gap between MiniMax’s speed claim and Artificial Analysis’s figure. Until then, the confirmed picture is a low listed token price, strong but not leading index performance, and important testing and release details still outstanding.

Key Questions

When did MiniMax release M3.1 Flash?

MiniMax announced the release on October 19, 2026. The model was available through the MiniMax API, M Plan subscription, MiniMax Code and OpenRouter, according to the review.

How much does M3.1 Flash cost?

The listed rate is $0.15 per million input tokens and $0.60 per million output tokens, with cached input at $0.015 per million tokens. These are token prices, not guaranteed costs for a complete task.

Did M3.1 Flash beat the leading open models?

Not on the Artificial Analysis comparison cited in the review: M3.1 Flash scored 43, below MiMo-V2.6-Pro at 46 and DeepSeek V4.1 Flash 1017 and GLM-5.3 at 45. The review also reports mixed results on MiniMax’s own benchmarks, which were self-reported.

What did the 24-article newsroom test find?

The review tested 12 real news sources in English and German, producing 24 articles judged for source fidelity, invented facts, clarity, usefulness, structure and polish. The supplied source excerpt does not include the scores, so the test’s outcome cannot be reported from the available information.

When are the model weights expected?

MiniMax said the weights would arrive on October 23, 2026. The source material does not confirm whether the release took place or state what license would apply.

Source: ThorstenMeyerAI.com