Published: 28 September 2026
MiniMax released M3.1-Flash-Preview on 27 September 2026, a coding model that reads up to 1,000,000 tokens and runs only inside the company’s own MiniMax Code tool. It is the first M3.1-branded model anyone can use. MiniMax published no price, no speed figure and no benchmark, for a model whose name is Flash.
What is MiniMax M3.1-Flash-Preview?
M3.1-Flash-Preview is aimed at everyday development work rather than hard problems: bug fixes, small features, the routine end of a coding queue. MiniMax describes it as fast, reliable and ready for real work, and that is close to the whole of what the company said about it.
The one genuinely documented feature is the effort setting. You choose how hard the model thinks from five levels: low, medium, high, xhigh and max. Leave it out and you get max. Unlike MiniMax M3, you cannot turn thinking off at all; asking for that returns an error. So the five levels have a floor, not an off switch.
How you actually reach it:
- Available only inside MiniMax Code, not as a standalone API
- Reported to need a Token Plan subscription rather than pay-as-you-go
- Token Plan tiers run $22, $55 and $132 per month
- Context window of 1,000,000 tokens, confirmed in MiniMax’s own documentation
- No open weights, no model card, no parameter count
- A double-points promotion runs from 28 September to 7 October 2026
How does M3.1-Flash-Preview compare to MiniMax M3?
Take this from the table: the useful comparison is not between MiniMax and its rivals, it is between what MiniMax published this time and what it published last time.
| M3.1-Flash-Preview | MiniMax M3, the shipped sibling | |
|---|---|---|
| Context window | 1,000,000 tokens | 1,000,000 tokens |
| Price per million input / output tokens | not published | $0.30 / $1.20 |
| Speed | not published | about 100 tokens per second |
| SWE-bench Verified (real bug fixes) | not published | 80.5% |
| Where you can use it | MiniMax Code only | API and MiniMax Code |
| Open weights | No | No |
The M3 figures are MiniMax’s own, from its documentation and launch materials. The M3.1-Flash-Preview column is not an omission on our part; those fields have no published value anywhere.
What MiniMax is not saying
Almost everything. There is no price per token, no tokens per second, no benchmark result on any public test, no parameter count and no model card. MiniMax quotes speeds for its other models in the same documentation, which makes the silence on this one conspicuous rather than accidental.
The name is the claim, and the claim is unsupported. Calling something Flash sets an expectation about latency that MiniMax has not measured in public, and the model is absent from third-party speed trackers because nobody outside MiniMax Code can call it. A model you can only reach through one vendor’s own tool cannot be independently benchmarked at all.
What this means
Safe to ignore unless you already pay for MiniMax Code. There is nothing here to evaluate. A preview with no price cannot be budgeted, a model with no benchmark cannot be compared, and a model locked inside one tool cannot be put behind your own router alongside the models you already use. If you are shopping for a cheap coding model today, MiniMax M3 is the one with a price and a published 80.5% on SWE-bench Verified.
Worth watching for what it signals rather than what it does. The mandatory thinking floor is a real design decision, not a relabel, and MiniMax has a track record of shipping the full model at an aggressive price a few weeks after a preview. The moment to look again is when M3.1 gets a price and an API. Until then this is a product announcement wearing a model release’s clothes.
For more information, visit the official announcement of M3.1-Flash-Preview in the MiniMax documentation.
