MiMo-V2.6-Pro vs MiMo-V2.6: When a Model Name Stops Being a Model

Artificial Analysis has a page for MiMo-V2.5. It has no page for MiMo-V2.6. That absence is the whole story of this matchup: last generation, the bare version number pointed at a checkpoint you could download and call. This generation it does not. Xiaomi MiMo-V2.6-Pro is a model with weights, a price and an independently measured score. "MiMo-V2.6" is a label that sits above it, and the two are not in the same category of thing — which makes this a comparison between a checkpoint and a name, not between two models.

That is worth being precise about, because the difference has a price attached. Xiaomi MiMo-V2.6-Pro lists at $0.435 per million input tokens and $0.87 per million output. Xiaomi MiMo-V2.6-Flash, the other checkpoint the label covers, lists at $0.14 and $0.28. If you picked the wrong one of those because a catalogue, a config file or a procurement sheet said "MiMo-V2.6" and nothing more, you are off by a factor of three before you have made a single call.

What "MiMo-V2.6" resolves to

Three different things, depending on where you read it.

The first is the series. Xiaomi's own release material describes the V2.6 generation as containing two native omnimodal checkpoints, Pro and Flash, published as separate training runs. The Hugging Face collection that holds them is named, plainly, MiMo-V2.6 — it currently lists three items, counting the Pro-UltraSpeed serving tier, and that collection name is the single most likely place for someone to learn the bare version number and assume it names a model.

The second is shorthand for the flagship. In most launch coverage and in most shorthand conversation, "MiMo-V2.6" means the trillion-parameter tier, because that is the model the coverage was about. This is the reading that produces the least damage, since it usually resolves to the model people meant.

The third is the one that costs money: a catalogue or config where "MiMo-V2.6" has collapsed onto whichever checkpoint that vendor happens to serve. Repositories are named MiMo-V2.6-Pro-RL and MiMo-V2.6-Flash-RL, and the -RL suffix is post-training lineage rather than an adapter sitting on a base model — a detail worth knowing before you go looking for the non-RL weights that pair with them, because there are none.

The mid-tier that stopped existing

The ambiguity is new, and the reason is that the middle of the lineup disappeared.

Artificial Analysis still lists a model called MiMo-V2.5, released 22 April 2026, scoring 25.17 on the index at $0.14 per million input tokens and $0.28 output, with open weights. It sits below MiMo-V2.5-Pro, which scored 26.0 on the same scale at $0.435 and $0.87 with 1M context and 1023B total parameters. That was a three-rung generation: a cheap bare-name model, a flagship, and nothing forcing you to choose between the two extremes.

The V2.6 generation has two rungs that anyone can call: Pro at the top and Flash at the bottom. There is no mid-tier checkpoint to absorb the traffic that falls between them, and none of the launch material describes one. Xiaomi did publish a MiMo-V2.6-Distill-Qwen-9B starting point and an RL framework alongside the weights, but a distillation starting point for training work is not a hosted mid-tier for inference.

Practically, that means the decision this page describes — Pro or the bare name — is really a decision about which rung you are standing on. If what you actually want is the old MiMo-V2.5 slot, the closest thing in the new generation is Flash, at the same $0.14/$0.28 and with a much higher capability claim attached.

The upgrade the name hides

Strip the naming question away and there is a clean, large, well-evidenced fact underneath it, which is that the flagship moved a long way in one generation without changing its price.

• Index score — MiMo-V2.5-Pro 26.0 on Artificial Analysis Intelligence Index, against 46.3 for Xiaomi MiMo-V2.6-Pro, both as measured by Artificial Analysis rather than by Xiaomi
• Per-token price — unchanged: $0.435 in, $0.87 out, on both generations
• Cost per index task — $0.054 for MiMo-V2.5-Pro against $0.1332 for MiMo-V2.6-Pro, the increase coming from verbosity rather than from a rate change
• Parameters — 1023B total and 42B active, against 1.02T total and 42B active
• Context — 1M tokens, unchanged
• Licence — MIT, unchanged
• Release date — 22 April 2026, against 21 September 2026

The score is the part that matters and it is the part the naming obscures. Twenty index points in five months, on an unchanged rate card, is a larger generational step than the release coverage framed — most of it went to "top open-weights model", which is a claim about the field rather than about the upgrade. If you are running the previous flagship and wondering whether a migration is worth the engineering, that single line is the answer, and it is a stronger argument than anything in the marketing material.

The caveat is the same one that attaches to every number from Xiaomi in this release. The 26.0 and the 46.3 are both Artificial Analysis measurements, which makes them comparable to each other and to other models on the same index version. The vendor's own benchmark tables — DeepSWE v1.1, Terminal Bench, AutomationBench and the rest — are not, and they should not be used to compute an upgrade delta. Mixing a measured score with a vendor score to produce a "points gained" figure is the most common error in coverage of this release.

Why a name is an operational problem

None of this is pedantry once the label reaches a system. A model identifier that resolves to two checkpoints with a 3.1x price difference is a billing incident waiting for a trigger, and there are three places it reliably goes wrong. Configuration and environment files, where a string like mimo-v2.6 gets written once and never revisited. Cost sheets, where a per-token estimate gets attached to the series name and then applied to the wrong tier. And evaluation rows, where a result gets filed under the family and later read as a property of the flagship.

There is a fourth place, and it is the one that bit this release specifically: the parameter counts. Xiaomi MiMo-V2.6-Pro is documented at 1.02T total and 42B active, and its Hugging Face repository reports a Safetensors model size of 524B. MiMo-V2.6-Flash is documented at roughly 309B total and 15B active, and its repository reports 159B. Neither pair reconciles. If your capacity planning reads the model card and your provisioning reads the repository, you will size the same deployment two different ways.

Pin the checkpoint name everywhere, not the series name. That is the whole remedy, and it costs nothing until the day it saves a quarter's budget.

How this resolves on a router

The naming problem is a catalogue problem, and a catalogue is the one place it is easy to fix. Addressing models by precise identifier through one API key, rather than by family name across several provider accounts, means the string in your config is the same string the billing system sees — which is the property a name like "MiMo-V2.6" destroys. That is also why we pass provider list price through at 0% markup rather than applying our own margin: the price on the page is the price the provider set, so a tier you misidentified is wrong in the same direction for everyone rather than wrong in a direction specific to us.

To be plain about the limits: we do not host either MiMo-V2.6 checkpoint, so neither is routable through us today, and neither is deployed by any third-party inference provider according to its own Hugging Face repository. The route to both is Xiaomi's own platform and the catalogues that carry it. But the identifier discipline this page is arguing for applies to every model you call, and it is the reason a routing layer is worth having even for the models you are not routing yet.

What to take away

If you came here to compare two models, the comparison is not available — one of these things is not a model. Xiaomi MiMo-V2.6-Pro is the checkpoint: 46.3 on the index, $0.435 and $0.87 per million tokens, MIT weights, 1M context, and the only open-weights entry on Artificial Analysis's cost Pareto line above the near-free tier. MiMo-V2.6 is the label above it and the label beside it, covering a cheaper checkpoint at one third the price and a faster serving tier at ten times the price.

The comparison worth making instead is the one the naming hides: this generation's flagship against the last one that carried a bare version number. Twenty index points, five months, unchanged rates. That is the number to put in front of whoever signs off on the migration — not the one about being first on the leaderboard.