MiniMax M3.1 Flash: What Its Newsroom Performance Says About Open Models

🔍 Read the full analysis: MiniMax M3.1 Flash: What Its Newsroom Performance Says About Open Models on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

• Business-only prices and quantity discounts
• Tax-exempt purchasing
• Multiple users, one account, clear invoices

Create a free Business account →Free to register

As an affiliate, we earn on qualifying purchases.Hide

MiniMax released M3.1 Flash on October 19 at $0.15 per million input tokens and $0.60 per million output tokens, pitching it as a fast, inexpensive model for agent tasks. A review published by ThorstenMeyerAI.com on October 20 found a mixed performance picture: an independent index measured lower speed than MiniMax claimed, and the review’s own newsroom test covered 24 articles but its scores are not present in the supplied source material.

The model is available through the MiniMax API, M Plan subscription, MiniMax Code and OpenRouter. OpenRouter offered free access for the first 72 hours after release. The listed price for cached input is $0.015 per million tokens. The source says the new model costs half as much per token as MiniMax M3; MiniMax also claims it costs about 70% less per agent task. Those are different comparisons: the per-task figure is the company’s claim, not an independently established result.

MiniMax advertises a one-million-token context window and says everyday generation speed exceeds 200 tokens per second, with a median above 170. The company describes M3.1 Flash as “nearly three times faster” than M3 and attributes its design to “more aggressive sparsity.” Artificial Analysis, which evaluates models across a shared set of tests, measured 120 tokens per second and 1.08 seconds to first token. That measurement is below MiniMax’s stated speed, though the source does not establish that the tests used identical settings.

The review says MiniMax’s published benchmarks are self-reported and that its comparison includes older versions of some competing models. In the company’s table, DeepSeek V4.1 Flash 1017 scores higher on DeepSWE and Terminal-Bench, while M3.1 Flash leads Automation-Bench by 1.7 points. The review also reports that Artificial Analysis gave M3.1 Flash an Intelligence Index score of 43, below several listed open-weight models. Artificial Analysis classifies it as proprietary because its weights have not yet been released.

Price Meets Real-World Performance

M3.1 Flash’s headline appeal is its low per-token price, but the review highlights why price alone does not determine operating cost. Artificial Analysis says the model used 180 million tokens to complete its index evaluations, compared with a median of 99 million. That is a raw token-use comparison in that evaluation, not a universal measure of how much the model would consume in other workloads. It does suggest that buyers should compare the cost of completing a task, not just the rate charged for each token.

For teams using models to draft or process news, source fidelity and invented facts can matter as much as speed. The source describes a test on 12 real news sources, each turned into English and German articles, then scored by two separate model-based judges. However, the supplied material ends before giving the results table. It is not possible from this source to say how M3.1 Flash scored, or whether it performed better or worse than the other models tested.

The release also matters to developers comparing open models with proprietary services. The promised weights could make M3.1 Flash available for uses beyond MiniMax’s hosted services, but the model is not open-weight at the time described. A public release date, without the weights or license yet in hand, does not settle what developers will be allowed to do or how the model can be run.

From Space Bunny to M3.1 Flash

Before the named release, the model appeared anonymously on OpenRouter as Space Bunny Alpha. According to the review, it went live on September 23, topped OpenRouter’s usage ranking for 10 days, and was removed on October 5. MiniMax says it processed 56.7 trillion tokens on OpenRouter during that period and more than 200 trillion tokens overall. Those usage figures are attributed to MiniMax and are not independently verified in the supplied material.

The review says the anonymous provider could retain prompts during the Space Bunny period, although it was not allowed to train on them. It reports that MiniMax has not explained what happened to that data. That question may matter to people and organizations that used the endpoint before the provider’s identity was disclosed.

MiniMax says the M3.1 Flash weights will arrive on October 23. As of the review’s October 20 publication, it had not announced a license or official parameter count. The source mentions an outside report of 428 billion total parameters and 23 billion active parameters, but says MiniMax has not confirmed those figures; they should not be treated as established specifications.

““nearly three times faster” than M3”

— MiniMax

Weights, Data and Test Scores

Several points remain unresolved. The promised October 23 weight release had not happened by the October 20 review, and no license or company-confirmed parameter count was available. The source also does not include the newsroom benchmark’s score table, despite describing the 24-article test and its judging method. Its results cannot be summarized or compared from the material provided.

MiniMax’s benchmark and security results, including its reported 82% score on CyberGym, are described by the review as internal or self-reported evaluations that had not been independently reproduced. The review also notes that MiniMax’s “first among Chinese models” claim for Agents’ Last Exam does not include DeepSeek in that row, while DeepSeek’s model card reports a different score that may use another setup. The two figures are not directly comparable based on the information supplied.

It is also unclear how MiniMax handled prompts retained during the anonymous Space Bunny period. The source gives no account of the data’s current status. Finally, speed and cost can vary with model settings, task type and evaluation method; the material does not provide enough detail to reconcile the company’s speed claim with the independent measurement or to generalize the newsroom test to all publishing work.

The October 23 Weight Release

The next stated milestone is MiniMax’s planned release of M3.1 Flash’s weights on October 23. Developers and researchers will be able to assess the files and license if the release proceeds, but the source does not confirm that either will be available on that date. Independent tests after release could also clarify how the model performs outside MiniMax’s own benchmark selection.

For newsroom performance, the immediate missing information is the score table for the 24-article evaluation, including the results for English and German and comparisons with the other models. The review says two judges scored source fidelity, fabricated facts, clarity, usefulness, structure and polish, but the supplied excerpt does not state those scores. Until they are available, claims about M3.1 Flash’s newsroom quality should remain open.

Users who tried Space Bunny Alpha may also look for an explanation from MiniMax about the prompts retained during that period. The review reports no answer so far. The release, its license, independent replication of reported benchmarks and clarification of prompt handling are the developments that could materially change the current assessment.

Key Questions

When did MiniMax M3.1 Flash launch?

MiniMax released M3.1 Flash on October 19, 2026, according to the October 20 review. It was listed on the MiniMax API, M Plan subscription, MiniMax Code and OpenRouter.

How much does M3.1 Flash cost?

The listed rate is $0.15 per million input tokens, $0.60 per million output tokens and $0.015 per million cached input tokens. MiniMax separately claims its cost per agent task is about 70% lower than M3; the source does not independently verify that claim.

Are the model weights available?

Not at the time of the October 20 review. MiniMax said the weights would arrive on October 23, but the supplied material does not confirm their release or specify a license.

How did M3.1 Flash perform in the newsroom test?

The review says it tested the model on 24 articles: 12 sources written in both English and German, judged by GPT-6.1 Sol and Claude Opus 5.5. The supplied source omits the results table, so its scores and comparative ranking cannot be reported.

What remains unclear about the earlier Space Bunny endpoint?

The review says the anonymous provider could retain prompts but could not use them for training. It also says MiniMax has not explained what happened to that retained data.

Source: ThorstenMeyerAI.com