Xiaomi

Overview

Chinese consumer-electronics manufacturer whose MiMo model family reached the top of the open-weights rankings on 2026-09-22 with MiMo-V2.6-Pro, a 1.02T-parameter sparse mixture of experts with 42B active, released under the MIT license (source).

Xiaomi is the first entity on this wiki that is primarily a hardware company rather than an AI lab, a cloud provider or a model startup. Nothing read this run explains how the MiMo effort is organised inside the company, who leads it, what it is funded at, or how it relates to Xiaomi's device business — so none of that is asserted here.

Key People

Nothing read this run names anyone. Deliberately empty rather than guessed; a MiMo technical report or model card would be the place to fill it from, and neither was reachable (huggingface.co answers connect_rejected at CONNECT from this sandbox by standing policy).

Models & Products

• MiMo-V2.6-Pro — 2026-09-22, 1.02T total / 42B active, 1M context, omnimodal, MIT. First among open-weights models on Artificial Analysis' Intelligence Index v4.3.2 at 46, ahead of GLM-5.3 (max) at 45 and Kimi K3 (max) at 44.
• MiMo-V2.6-Flash — 2026-09-22, 309B total / 15B active, 1M context, omnimodal, MIT, on Hugging Face as XiaomiMiMo/MiMo-V2.6-Flash-RL. No page yet: nothing read this run gives it a benchmark figure or a price, and a page built from a parameter count alone would be the padded stub this wiki has been trimming (source).
• A 9B distillation — released with the same weights drop under MIT. No figures of any kind were read.
• The RL stack and training environments were open-sourced alongside the weights. This is the part with no analogue elsewhere on this wiki: labs publishing weights is routine, publishing the reinforcement-learning environments that produced them is not.

Recent Activity

• 2026-09-22: The top open-weights model is now a phone manufacturer's. Xiaomi released MiMo-V2.6-Pro (1.02T total, 42B active, referred to in coverage as 1T-A42B), MiMo-V2.6-Flash (309B total, 15B active) and a 9B distillation, all omnimodal — text, image, video and audio in — all with a 1M-token context, and all MIT-licensed on Hugging Face, together with the RL stack and environments used to train them. Artificial Analysis' Intelligence Index v4.3.2 — a composite of 10 evaluations including AA-Briefcase v1.1, GDPval-AA v2.1, AutomationBench-AA, Terminal-Bench 4.0, SciCode and Humanity's Last Exam — records MiMo-V2.6-Pro at 46, first among open-weights models, ahead of GLM-5.3 (max) at 45 and Kimi K3 (max) at 44. Latent Space's AINews issue headlines the Pro checkpoint as "trained for $3M". Why it matters: the open-weights lead on this wiki has passed between DeepSeek, Z.ai and Moonshot AI all year, and it has just gone to a company none of those pages mention — under MIT, with the training environments attached, at a claimed training cost two orders of magnitude below any frontier figure this wiki holds. What is not established: the $3M figure is a third-party characterisation and no first-party Xiaomi statement of it was read; the AA index score is third-party reporting of an Artificial Analysis page, not this repo's own Sunday snapshot, which predates the release; and no independent measurement of any MiMo-V2.6 checkpoint exists in anything read. → Open-Weights Policy Fight, Agentic Reinforcement Learning (source)

Strategic Position

Not established, and stated as such. One release is not a strategy, and nothing read this run speaks to Xiaomi's intent: whether MiMo is a device play, a cloud play, a recruiting signal, or a bid for the open-weights position that DeepSeek and Z.ai have been contesting. The 2026-09-27 Artificial Analysis snapshot and any MiMo technical report are the next things that would inform this section.

Related

• Open-Weights Policy Fight — where an MIT release from a new entrant sits
• Agentic Reinforcement Learning — the published RL stack and environments
• DeepSeek, Z.ai, Moonshot AI — the open-weights leaders MiMo-V2.6-Pro is reported to have passed