The MiniMax Family Explained: M3 and Native Multimodality

MiniMax M3, released 1 June 2026, occupies a position no other model in the open-weight field quite matches: a million-token context window that is natively multimodal rather than text-only with an image adapter bolted on.

It is also the model whose published pricing is least settled. Sources disagree, materially, and that is worth knowing before you build a cost model on a number you read once.

Native multimodality is a different thing from vision support

Many models can accept an image. Fewer are trained from the start on interleaved modalities, and the difference shows up in how well the model reasons across them rather than about them separately.

A model with a bolted-on vision path tends to describe an image accurately and then reason about the description. A natively multimodal model is more likely to hold the image and the text in the same reasoning process — which matters when the answer depends on a relationship between something in the picture and something in the prose.

The practical test is not "can it read this screenshot" but "can it answer a question that requires combining the screenshot with the surrounding text". That is where the architectures separate. How multimodal models work covers the mechanism.

A million tokens that can hold pictures

Combine native multimodality with a 1M context window and you get a capability that is genuinely awkward to assemble any other way: an enormous input in which images and text are interleaved.

Concrete cases where that is the whole requirement include reviewing a long design document with embedded mockups, debugging from a stack trace alongside screenshots of the failing interface, and processing scanned or image-heavy documentation where the layout carries meaning that plain text extraction destroys.

Without native multimodality you build a pipeline: extract images, caption them separately, splice the captions back into the text, and hope nothing important was lost in the captioning. That pipeline is real work and a real source of silent errors. The 1M-context comparison covers who else ships the window, though M3 is alone on the multimodal axis.

Read the pricing carefully

This is the part that requires care. Published pricing for M3 differs between sources — one set of figures is roughly half the other, and the discrepancy is large enough to change a build-or-buy decision.

We are not going to pick one. Quoting a confident number here would be worse than useless if you budget against it and it is the wrong one. Check MiniMax's own pricing page for the tier and region you will actually use, and check it again before you commit, because rates move.

The more general lesson is worth internalising: third-party pricing tables are frequently stale, frequently omit tiering, and frequently do not distinguish cached from uncached input. For anything where cost is a deciding factor, the vendor page is the only source worth trusting, and your own measured invoice is better still. Input against output token pricing covers why headline rates mislead even when they are accurate.

The benchmark position

M3 reports 59.0 percent on SWE-bench Pro, and that figure is vendor-claimed rather than independently reproduced at the time of writing. Its Artificial Analysis Intelligence Index sits around 44.

Two things follow. First, the SWE-bench figure is a Pro score, and Pro is the harder variant — placing it in a column next to Verified scores from other models produces a comparison that means nothing. Second, vendor-claimed is not the same as wrong, but it warrants the same caution you would apply to any self-reported result.

Against the field, an index around 44 puts M3 roughly alongside DeepSeek V4 Pro and below GLM-5.2 at around 51 and Kimi K3 at around 57. If pure text reasoning is what you need, the ranking is unfavourable. If you need reasoning over interleaved images and text at length, the ranking is beside the point because the alternatives do not offer the capability. Why benchmarks disagree covers how much weight any single index deserves.

What about M2.7

MiniMax M2.7 exists and predates M3. Its specifications are not well established in public sources, and we are not going to guess at them — a parameter count or context window invented for the sake of completeness is worse than an admission of uncertainty.

If you are evaluating it, treat the vendor documentation as the only authority, and be alert to the possibility that community-maintained comparison tables have carried forward figures from a different release. This happens more than you would expect across the whole field.

The practical position: build against M3, which is documented, and treat older tiers as something to verify rather than assume. Record the exact model identifier and the figures you relied on at the moment you make a decision.

When M3 is the right choice

Choose it when your inputs genuinely mix images and text at length, and the relationship between them matters. That is the case where it is not competing with the rest of the field — it is the only convenient option.

Choose it also when you would otherwise build a captioning pipeline to fake multimodality, because that pipeline is more expensive to maintain than it looks and loses information in ways that are hard to detect.

Look elsewhere for pure text reasoning, where the index positions are unfavourable, and for cost-driven long-context reading, where a cheap text model with a 1M window reads the same input for less. The comparison against DeepSeek V4 Flash works through that second case, and the M3 guide goes deeper on setup.