MiMo-V2.6-Pro: Xiaomi's Trillion-Parameter Flagship, Read From the Card

MiMo-V2.6-Pro is the flagship checkpoint of the MiMo-V2.6 series from the X​iaomi MiMo team: a sparse mixture-of-experts model with roughly 1.02 trillion parameters in total and 42 billion active per token, native text, image, video and audio input, a 1M-token context and MIT-licensed weights. It sits above MiMo-V2.6-Flash, the 309B efficiency checkpoint, in a series whose parts are separate training runs rather than one model at two sizes. Readers arrive at this name spelled three ways — mimo-v2.6-pro as the platform identifier, mimo v2.6 pro with spaces, and mimo v2 6 pro with the version split into digits — and all three point at the same checkpoint, the one whose weights live at X​iaomiMiMo/MiMo-V2.6-Pro-RL.

This page is the reference for the model itself. The comparisons and the cost arithmetic already exist on this blog and are indexed below rather than repeated; what follows is what the model is, what it publishes about itself, and which figures on the card are worth trusting as far as their source allows.

Identity: the lab, the class, and what "Pro" distinguishes

The publisher is X​iaomi, through the group that signs its technical reports as the X​iaomi MiMo Team. The weights, the platform documentation and the API all come from that one organisation, which matters for provenance: every capacity figure below is the vendor describing its own checkpoint, and none of it has been reproduced here.

"Pro" is a position inside a three-model series, not a quality adjective. X​iaomi's own series note calls Pro the most powerful flagship reasoning model to date — omni-modal, ultra-high performance, trillion-parameter — built for complex projects, long-horizon tasks, high-stakes work and research; Flash is the full-modality, low-cost reasoning model, and Pro-UltraSpeed is a serving tier that carries the same checkpoint at up to twenty times the output speed for latency-sensitive work. The platform's quick-selection guide routes complex projects, long-term tasks and high-value work to Pro, high-frequency professional office calls to Flash, and anything sensitive to response speed to Pro-UltraSpeed. Three published positions, one series — not three sizes of one training run.

The version carry is worth a line, because it is where returning readers lose their footing. MiMo-V2.5 Pro, the previous flagship, is scheduled for retirement: X​iaomi's platform gives a deprecation time of 10:00 Beijing time on 21 October 2026 for both mimo-v2.5 and mimo-v2.5-pro, and recommends switching to the new version. Pro is that new version, so a search for "MiMo Pro" that lands on 2025 material is landing on the model this one replaced.

The specification, taken from the model card

The card's own summary block is the shortest honest description of the model, and every line of it is the vendor's:

• Parameters — 1.02T in total, 42B active per token, a sparse mixture of experts. Repository metadata counts 1,024,216,603,392 parameters across the published shards.

• Context and output — a 1M-token context window and up to 128,000 tokens of output.

• Modalities — text, image, video and audio input in one model, with a 681M-parameter MiMo ViT for vision and a 308M AudioTokenizer plus a 127M audio patch encoder for audio.

• Backbone — 70 layers, 60 sliding-window and 10 global; hidden size 6,144; 384 routed experts with 8 activated; sliding window 128. The first block runs global attention with a dense FFN, and no block uses shared experts.

• Speculative decoding — a 5-layer sliding-window multi-token-prediction drafter, which the card describes as predicting seven further tokens per forward pass.

Two notes on reading that list. The parameter total is a vendor figure, but it is the one an outsider can partly check, because the shard sizes are published in the repository's file listing; the activated figure has no such cross-check and should be taken as claimed. And the last two items sit slightly at odds with each other — a drafter described by its layer count and a drafter described by how many tokens it predicts per pass are two different measurements of the same component, and only the layer count appears in the architecture table.

Pro against base and Flash, in one place

These are the distinctions readers actually arrive with, both sides read from the same vendor documentation the same day:

• Scale — Pro is 1.02T parameters in total with 42B active; Flash is a separate training run at 309B total with 15B active.

• Depth — Pro runs 70 layers with 384 routed experts and 8 activated; Flash runs 48 layers with 256 routed experts and 8 activated.

• Price — Pro is US$0.435 per million input tokens and US$0.87 per million output tokens on a cache miss, with a cache hit at US$0.0036; Flash is US$0.14 and US$0.28, with a cache hit at US$0.0028.

• Selection — X​iaomi routes complex, long-horizon and high-value work to Pro and high-frequency professional office traffic to Flash.

• Shared — the 1M context window, the 128,000-token output ceiling and the MIT licence are identical across both.

The two comparisons this blog already runs go deeper than that list and are not reproduced here. The Pro-against-the-bare-series-name piece works through why "MiMo-V2.6" is a collection with more than one member rather than a third checkpoint sitting between Pro and Flash. The Pro-against-Flash piece carries the full architecture contrast, layer by layer, for a reader choosing between the two on a real workload.

Weights and licence for this build

There is one repository for this checkpoint, X​iaomiMiMo/MiMo-V2.6-Pro-RL, and its metadata is unambiguous: MIT licence, ungated, created 21 September 2026, last modified 9 October 2026, with a little over 96,000 downloads recorded at the time of writing. The parameters split across storage dtypes as roughly 1,000.2 billion in U8, 13.4 billion in FP8 (E4M3), 10.6 billion in BF16 and a trace in F32 — the shape of a checkpoint that ships its expert weights in a low-precision format and keeps the rest at BF16. The shards are large: a single expert-parallel file runs to about 4.15 GB, and the full download is on the order of several hundred gigabytes.

The -RL suffix is a lineage marker, not an adapter and not a quantisation. Both V2.6 checkpoints carry it, and the card positions the whole series around reinforcement learning at scale — one mixed run across coding, general agents, visual work and cybersecurity rather than separate per-domain runs, with a groupwise grading stage that ranks passing trajectories instead of scoring them pass or fail. None of that changes what the weights are; it explains why the repository is named the way it is.

Running it: what is actually published

The card gives two deployment paths and no more. SGLang is served with the ordinary python3 -m sglang.launch_server invocation against the repository path, and the card's own example uses a two-node layout with expert parallelism. vLLM is served with vllm serve plus the flags that name this model's parsers, --reasoning-parser mimo and --tool-call-parser mimo, with auto tool choice enabled and the generation config read from the model directory. Sampling on the card is temperature 1.0 and top_p 0.95 — warm settings, and a fair warning that the 0.7 many harnesses default to is not what the vendor tested.

Alongside the self-hosted routes, X​iaomi's platform publishes the operational envelope for its hosted endpoint: the 1M context window, a 128,000-token maximum output, and a rate quota of 100 requests and 10 million tokens per minute. The platform also lists a Batch API covering this model for non-real-time work at a lower rate. What is not published is any throughput or memory figure from X​iaomi itself, so this page does not carry one — the speed numbers that circulate come from third parties, measure their hardware rather than yours, and belong with their measurement conditions attached.

The scores, and who measured them

X​iaomi's card carries a benchmark table, and everything in it is vendor-reported: DeepSWE v1.1 at 71.9, Terminal Bench 2.1 at 89.9, Agents' Last Exam at 31.6, OSWorld-Verified at 82.0, GDPval-AA 2.1 at 1673, AutomationBench v1.0.6 at 53.1, and on the cybersecurity rows CyberGym at 94.0 and the vendor's own MiMo Cyber Bench at 80.2. Those are the numbers to quote when the question is what the lab claims. The same table prints figures for Flash, for the previous MiMo-V2.5 Pro and for three closed models, which is a reminder worth having: a vendor table's comparison rows are the vendor's selection, run on the vendor's harness choices.

The independent figure is a different thing, and it is the one that has moved since the launch coverage. Artificial Analysis scores MiMo-V2.6-Pro at 46 on its Intelligence Index, measured on version 4.3.2 of that index — the highest score that page records for a model with open weights, one point above GLM-5.3, two above Kimi K3, and behind the closed leaders it also evaluates. The same page prices a full index run at about US$206.66, records a 99% cache discount against list, and gives a cost per index task of roughly US$0.13. Read the two sources together and the shape of the model is legible without either being promoted: a vendor whose own table is generous, and a third party whose single index still puts it at the front of the open-weights field.

The cost work, indexed rather than repeated

Three published pieces already carry the price-gap and value argument for this model, and each answers a different question:

• The Pareto-frontier piece asks where MiMo-V2.6-Pro lands when cost per unit of measured capability is recomputed across the wider field rather than against one rival.

• The per-task cost-gap piece asks what the same measured capability costs when the cheaper side is the one with open weights.

• The same-score piece takes the case where two checkpoints finish within a fraction of a point on one index and the bill for running that index against each differs by more than an order of magnitude.

What belongs here instead is the price as a fact about the model, not as an argument. X​iaomi publishes two rate cards, and they are two published numbers rather than one converted into the other: the mainland card lists ¥0.025 per million input tokens on a cache hit, ¥3.00 on a miss and ¥6.00 for output, while the overseas card lists US$0.0036, US$0.435 and US$0.87 for the same three lines. Cache writes are free for a limited promotional period on both. Whether a given workload's bill resembles the cache-hit row or the cache-miss row is a question about the workload, and the three pieces above are where that question gets argued.

One structural point outlives any single rate card: a published price is only useful if it is the price you are actually billed. That is why the catalogue behind this blog carries 200+ models at provider list price with 0% markup passed through, rather than a margin applied on top of whatever the provider set — a change the vendor makes is a change on your bill the same day, not at a renewal.

What is not settled

Two loose ends, neither of them a defect, both worth knowing before you quote the spec sheet. The card's speculative decoder is described in two different units — five layers in the architecture table, seven predicted tokens per pass in the prose — and only the first is stated as configuration. And the modality claim is wider on the card than third parties currently record for the served endpoint: the weights accept four input types, while an evaluation site measuring a hosted API may show a narrower surface. That is the ordinary gap between what a checkpoint can do and what a given endpoint exposes, and anyone planning audio or video work should confirm it against the endpoint they intend to call rather than against this page.

Reading the page back

The short version, for a reader who came from a search box and wants the answer rather than the argument: MiMo-V2.6-Pro is X​iaomi's trillion-parameter flagship — sparse mixture of experts at roughly 1.02T total and 42B active, native text, image, video and audio input, a 1M-token context with 128,000 tokens of output, MIT-licensed weights at X​iaomiMiMo/MiMo-V2.6-Pro-RL, 46 on the Intelligence Index where that index is the source and 71.9 on DeepSWE v1.1 where X​iaomi is, priced at US$0.435 and US$0.87 per million tokens on the overseas card with a cache hit at US$0.0036, and separated from MiMo-V2.6-Flash by scale, depth, price and intended workload rather than by generation.

One caveat on the whole page, since this is a reference and not a news item: every figure here is as published on 13 October 2026 by the sources named beside it, and vendor-reported numbers are labelled as vendor-reported where they appear. A rate card, a cache price, a context ceiling and a third-party index score are all things that change without an announcement, so anyone leaning on one of them for a decision should check that line against its source rather than against this page.