Sources policy
Our Sources, Testing and Update Policy
Four evidence labels, used consistently on every page, and the rules for updating and correcting them.
Why this page exists
MiniMax M3.1 Flash launched with a dense, contradictory information environment: vendor benchmarks, third-party scores, a free window that lasted days, a retired preview name, conflicting statements about open weights, and coverage in two languages with different levels of sourcing. A reader deserves to know where each number came from.
So every factual claim on this site carries one of four labels.
The four evidence labels
| Label | Meaning | How we treat it |
|---|---|---|
| Official | Stated by MiniMax in its own documentation or launch post | Reported as a claim, with the date. Vendor benchmarks are labelled as vendor-reported, never as independent. |
| Third-party | Measured or published by another party, such as Artificial Analysis | Attributed to the source, and we note whether that source measured the model itself or received vendor-supplied performance data. |
| Our testing | Reproduced on our own machine | Only used where we actually ran it, with the environment recorded. Everything else is explicitly not our testing. |
| Community | Forum threads, GitHub repositories, user reports | Linked, never used as sole support for a factual claim about the product’s behaviour. |
Where a page contains something we have not verified ourselves, it says so in the same section rather than in a footnote.
Primary sources
- MiniMax API documentation — the reference for model identifiers, base URLs, parameters, effort levels, caching, limits, integration guides and pricing. Where our summary differs from the docs, the docs win and we change ours.
- MiniMax launch post and GA page — for capability claims, benchmark tables and the Space Bunny statement.
- Artificial Analysis model pages — for third-party measurements, always with the note that M3.1 Flash’s performance and speed fields are flagged there as vendor-supplied.
- Platform and distributor listings (OpenRouter, Ollama, Vercel AI Gateway, cloudprice) — for availability, provider counts and free-window announcements.
Known conflicts we keep visible
| Question | Conflicting statements | Our position |
|---|---|---|
| Are the weights open? | Homepage calls it an “open-weight model”; the launch post says weights come later; the Hugging Face repository returns 404; Artificial Analysis records isOpenWeights = false | Treated as not released, with all four sources shown |
| Is the model “93% hallucinated”? | Chinese coverage quotes a hallucination figure without naming the metric | The sourced figure is Artificial Analysis’s Omniscience hallucination rate (92.81%, accuracy 56.2%); we always name the benchmark |
| How fast is it? | Vendor says P50 170+ TPS; Artificial Analysis displays a median of 104.7 t/s on long prompts | Both are reported side by side, with the note that the Artificial Analysis figures are vendor-sourced for this model |
| Which benchmark wins? | DeepSWE has documented task-quality problems (>20.3% of tasks flagged) | Reported as a rough signal, never as a ranking verdict |
Review dates and updates
- Every page shows a reviewed date.
- We re-check the pricing page, the FAQ, the free-window information and the open-weights status whenever MiniMax publishes a release note, because those are the facts that change fastest.
- Pages that depend on our own testing are re-verified when the underlying tooling changes.
Corrections
If a fact on this site is wrong, outdated or over-stated, tell us with the page URL, the claim, and a source if you have one. We correct factual errors and note material changes to the sourcing policy on the about page.
Commercial relationships
We are not affiliated with MiniMax and receive no access, discount or support from them. If this site ever earns affiliate revenue from a provider link, that relationship will be disclosed on the page where the link appears and in this policy.
Accuracy over completeness
We would rather leave a gap labelled than fill it with a plausible number. Where information was unavailable at review time — Google search results we could not read, third-party reproductions of the speed claims, the reasoning variant of a competitor model — the pages say so.
Sources and version notes
Checked October 21, 2026. We label official claims, third-party measurements, our own testing and community reports separately — see the sources policy for what each label means.