Xiaomi's MiMo-V2.6 Landed on Hugging Face Without a Word

A week ago, two model cards on a public training dashboard were marked "stopped," and the standing line about them was that no weights existed yet. On Monday, without a blog post, a press release, or even a price list, Xiaomi's MiMo team quietly uploaded both checkpoints to Hugging Face. MiMo-V2.6-Pro, a 1.02-trillion-parameter model, appeared at 15:39 UTC. MiMo-V2.6-Flash followed eighteen seconds later. No announcement thread went out until hours afterward. The release itself, not any marketing around it, is what changed.

Table of contents

• What actually shipped
• Where it lands on the leaderboard
• Two models built for different jobs
• The unusually quiet rollout
• What the MIT license actually changes
• Where this fits in the open-weight race
• Frequently asked questions

What actually shipped

OrcaRouter's technical write-up, published the day of the release, is precise about the timeline: MiMo-V2.6-Flash-RL went live on Hugging Face as an ungated, MIT-licensed checkpoint with a technical report attached and 65 weight shards in the index, and MiMo-V2.6-Pro-RL followed less than twenty seconds later. CellCog's account adds the rest of the day's trail: the API rows appeared on OpenRouter at 20:07 UTC, and Xiaomi's own announcement on X went out at 20:51 UTC, more than five hours after the files were already downloadable.

Both models are natively omnimodal, meaning they take text, image, video and audio as input in one architecture rather than bolting a vision or audio module onto a text model. Both carry a 1,048,576-token context window. Pro is a sparse mixture-of-experts model with 1.02 trillion total parameters and 42 billion active per token; Flash runs 309 billion total with 15 billion active. SiliconANGLE also reports a third variant, Pro-UltraSpeed, which Xiaomi says generates output up to 20 times faster than Pro at comparable quality.

Where it lands on the leaderboard

CellCog reports Artificial Analysis scored MiMo-V2.6-Pro at 46 on its Intelligence Index v4.3, putting it first among the 114 open-weight models the index tracks and level with xAI's Grok 4.7, which was released the same day. That score sits seven points behind Claude Fable 5.1 and GPT-6 Astra, the current frontier of closed models.

Datanorth's comparison sharpens the pricing side of that result: Grok 4.7 costs roughly $2 per million input tokens, while MiMo-V2.6-Pro's own API charges $0.435, undercutting even the newly discounted GPT-6 Sol on a per-token basis, and OpenRouter and other third-party hosts price it lower still. Matching a frontier lab's score at under a quarter of its price is the headline number here, though CellCog is careful to note the gap widens on the hardest tasks: on Xiaomi's own agent benchmarks, Pro sits close to Claude Opus 5 and GPT-5.6 Sol on general tasks, but clearly behind both on the toughest terminal and exploit-focused evaluations.

Two models built for different jobs

OrcaRouter's breakdown of the two configs shows they are not simply a big model and a small model of the same design. Flash uses 48 layers, 39 with sliding-window attention and 9 with global attention, a hidden size of 4096, 256 routed experts with 8 active per token, and no shared experts. That architecture favors throughput and cost over raw capability, which matches Xiaomi's own description of Flash as the "efficiency-balanced checkpoint" of the series, against Pro as the flagship.

The Hugging Face model card, cited in Xiaomi's technical documentation, frames the training approach around scaling reinforcement learning rather than model size alone: a single mixed RL run spanning coding, general agent tasks, visual reasoning and cybersecurity tasks together, rather than separate training runs per domain, on the reasoning that skills learned in one area transfer to harnesses the model never saw during training.

The unusually quiet rollout

The gap between the weights going live and Xiaomi saying anything about it is the detail most coverage keeps returning to. OrcaRouter points out that Xiaomi's own blog, as of the day after release, still showed December's MiMo-V2-Flash launch as its newest post. There was no V2.6 launch essay, no published price list for the new generation at the time of upload, and no callable model identifier Xiaomi had formally announced for either checkpoint when the files first appeared.

Whatever the reason, cautious internal sign-off, a staged rollout, or something more mundane, the effect is that a major open-weight release arrived first as files a researcher could download and only later as a story a journalist could write. That is a different order of operations than the coordinated, multi-outlet product launches most large labs run, and it means anyone tracking open models has to check Hugging Face and public dashboards directly rather than waiting for an announcement to know a release has happened.

What the MIT license actually changes

MIT is one of the least restrictive open-source licenses available: commercial use, modification and redistribution are all permitted with essentially no conditions beyond keeping the license notice. LLM Stats' entry for MiMo-V2.6-Flash confirms both checkpoints ship this way, with no revenue cap and no research-only clause, the kind of restriction some other open-weight releases attach.

The practical catch is hardware. Xiaomi publishes no official hardware guidance for self-hosting either model, according to Datanorth, though OrcaRouter notes SGLang and vLLM serving recipes are available on the model cards. A 1.02-trillion-parameter mixture-of-experts model needs multi-GPU serving regardless of license terms; only the 9B distillation Xiaomi also released is practical to run on a single card. The license removes the legal barrier to using the weights. It does not remove the compute cost of actually running the large ones.

Where this fits in the open-weight race

MiMo-V2.6 lands in a crowded month for open-weight releases. DeepSeek shipped DeepSeek-V4.1-Flash days earlier, and both companies are chasing the same target: closing the gap to closed frontier models while keeping inference cheap enough that developers actually switch. Xiaomi's own agent benchmark numbers, placing it near Claude Opus 5 and GPT-5.6 Sol on general tasks but behind them on the hardest ones, describe that gap with more precision than a single leaderboard score can.

What the quiet rollout adds to that picture is a reminder that the open-weight race is now moving fast enough that some labs are shipping ahead of their own communications teams. A model that was "stopped" on a public dashboard a week earlier is now a file anyone can download, and the announcement was almost an afterthought.

Frequently asked questions