MiniMax Launches M3.1-Flash-Preview: Six Reasoning Tiers and a Million-Token Context, Tested
According to Xinzhiyuan (新智元), an anonymous model called Space Bunny Alpha appeared on OpenRouter and OpenCode on September 23, and developers compared its tokenizer and speculated it came from MiniMax; four days later, on September 27, MiniMax launched M3.1-Flash-Preview, offering six adjustable reasoning tiers, a 1M-token context and text, image and video input. The report also documents test results covering bug fixes, web page generation, 3D and games.
What changed
According to Xinzhiyuan (新智元), on September 23 OpenRouter and the coding tool OpenCode simultaneously listed a time-limited free anonymous model called Space Bunny Alpha, which OpenRouter labeled a Flash model with fast inference speed, adjustable reasoning effort, a 1M-token context, and text, image and video input. The report says that on the same day developers compared the model's tokenizer and concluded its text tokenizer was identical to MiniMax M3, and the next day someone used 24 probe prompts to compare token counts, with results matching MiniMax models, so many developers speculated the anonymous model came from MiniMax. Four days later, on September 27, MiniMax officially launched M3.1-Flash-Preview, which the report describes as likewise offering six reasoning tiers, a 1M-token context and text, image and video input, consistent with the anonymous model's characteristics; however, as of the report's publication, neither MiniMax nor OpenRouter had publicly claimed Space Bunny Alpha, so the link between the anonymous model and M3.1-Flash-Preview remains a developer speculation rather than an official confirmation. The report also records interface information showing six reasoning tiers and a maximum 1M context when selecting M3.1-Flash-Preview in MiniMax Code.
What it means for developers
For developers, the key point is that the capability boundary of the "Flash tier" is being redrawn. The report says M3.1-Flash-Preview supports a 1M-token context and direct text, image and video input, while reasoning effort is adjustable across six tiers, meaning the same model can shift along the cost-versus-quality curve according to task difficulty instead of requiring routing across models. The test scenarios described cover code repair, web page generation, 3D visualization and game development, and the report says it fixed user-reported bugs in the real open-source projects Hono and marked in 2.8 minutes and 3.7 minutes respectively, with official regression tests passing; these are the media outlet's own test conclusions, not third-party reproductions. Note that token-usage comparisons mentioned in the report involve models from other vendors and represent only the platform observations at that time, so they should not be used as a basis for current model selection. On pricing, the report gives no citable pricing details, so refer to the vendor's pricing page and the KimiSeek pricing page.
Integration and evaluation
The following is advice and conditional analysis, not reported fact. First, if you rely on one-shot injection of long documents or large codebases, treat M3.1-Flash-Preview's 1M context as a hypothesis to verify: the premise is that your knowledge-base chunking strategy allows a single injection close to the upper limit, in which case you should compare the answer consistency of chunk-retrieval approaches against whole-context injection, using key-information recall on a fixed question set and citation-location accuracy as evaluation metrics rather than subjective fluency. Second, the six reasoning tiers are an adjustable parameter, and the decision point is the default tier; if your workload is mainly short queries and latency-sensitive, start with the default tier and raise it only on samples where multi-step reasoning fails, measuring the pass-rate gain of the same prompt set between adjacent tiers alongside the change in per-request latency and token consumption. Third, the multimodal input in the report means the real choice is whether to let the model read images/video directly instead of transcribing to text first: when your input includes hand-drawn wireframes, screenshots or screen recordings, compare direct multimodal input against an OCR-plus-text pipeline, using end-to-end task completion rate and manual rework rate as criteria. Fourth, code repair and long context recur throughout the report, so before integration build a regression baseline on your own real bug set, recording the official test pass rate after fixes and whether new defects are introduced. All of this should first go through small-traffic canary testing before deciding whether to replace your existing Flash-tier model.
Limits and open questions
To be clear: the version in this report, M3.1-Flash-Preview, is not currently listed in the KimiSeek live catalog, and the API availability mentioned points to the vendor MiniMax and its partner platform OpenRouter, not the KimiSeek platform. Second, the identity of Space Bunny Alpha and M3.1-Flash-Preview comes from developer comparisons of tokenizers and token counts, which is speculation; neither MiniMax nor OpenRouter has claimed it, so it cannot be treated as a confirmed conclusion. Third, results such as the pelican SVG, ticket-system fix, temperature spiral chart, SQL simplification fix, rocket and black hole 3D scenes, and the Night Harbor Operation game all come from the media outlet's own tests, and the test conditions, prompts and hardware environment were not fully disclosed, making them hard for third parties to reproduce; performance and latency figures should not be extrapolated directly. Finally, the model is in Preview, so interfaces, tier behavior and available regions may change, and the report provides no citable billing basis, so consult the vendor's pricing page and latest documentation before deployment.