Xiaomi Open-Sources MiMo 2.6 Pro and Flash With Omnimodal Inputs
The two models combine text, image, video and audio inputs with a million-token context window, while the vendor also releases training resources and RL code.
Edited by Tyronne Panaino
Xiaomi released and open-sourced the MiMo 2.6 series on September 22, pairing a larger Pro model with a Flash variant aimed at a different balance of capability, efficiency and cost. Both models accept text, images, video and audio, and Xiaomi lists a context window of up to one million tokens.
The release matters to model developers because it combines downloadable weights with more of the surrounding research and training stack. It also gives teams another open-model option for long agent traces, repository-scale work and multimodal inputs. Those use cases are the vendor's intended scope; this run did not independently reproduce the training results, benchmark table or product demonstrations.
Pro and Flash share an omnimodal design
Xiaomi's release article describes MiMo 2.6 Pro as its most capable model in the series and MiMo 2.6 Flash as the efficiency-oriented option. The company says both are available through its AI Studio, MiMo Code, desktop product and API platform, as well as OpenRouter. Availability on those services does not by itself establish identical limits, latency or pricing across every route.
The MiMo 2.6 Pro model card identifies a sparse mixture-of-experts architecture with about 1.02 trillion total parameters and 42 billion activated parameters. It lists a million-token context length, separate vision and audio encoders, and a five-layer speculative decoder. The card marks the model with an MIT license and provides download entries for both Pro RL and Flash RL checkpoints.
Those specifications make self-hosting a different decision from calling a managed endpoint. The total model scale and the published multi-GPU serving recipes imply substantial infrastructure requirements for the full Pro checkpoint. Teams evaluating local deployment should measure memory, throughput and multimodal preprocessing on their own hardware instead of assuming that an open license makes the operational footprint small.
Xiaomi opens more than the checkpoints
The release article says Xiaomi is also opening the full technical report, training environments and reinforcement-learning code. Its training description mixes coding, general-agent, visual and cybersecurity tasks across several harnesses, with an agentic grader comparing trajectories inside each group. This is Xiaomi's account of its training design, not independent evidence that the method transfers reliably to a buyer's workloads.
The company presents MiMo 2.6 as a model for software work, computer use, three-dimensional content, embodied simulation, design, media creation and research assistance. The breadth is notable, but the demonstrations remain curated first-party examples. A useful evaluation should separate basic artifact availability from task quality: first confirm that the intended checkpoint, code and license are accessible, then test a bounded workflow with reproducible inputs and failure criteria.
Evidence limits and next checkpoints
Xiaomi publishes benchmark scores, training-cost figures and a claim of much faster output from an UltraSpeed service. This article does not repeat those performance claims as established comparisons because the run did not fetch independent methodology or replication evidence. The most firmly supported change is the release itself: two omnimodal MiMo 2.6 variants, downloadable model artifacts, vendor-hosted access and named research resources.
The next verifiable checkpoints are independent evaluations of the released checkpoints, practical hardware profiles and confirmation that the published training environments and RL code reproduce the described workflow. Until those arrive, teams should treat capability, efficiency and speed comparisons as vendor-reported.
Status
Confirmed release; medium internal confidence. Xiaomi's official article and model card establish the artifacts, architecture summary, license and announced availability, but they do not independently validate benchmark, cost, speed or case-study outcomes.
Sources
Update note: Last reviewed 2026-09-26. We will revise this post if independent evaluations or reproducible deployment evidence materially change the picture.
Sources
- Xiaomi — Introducing MiMo-V2.6 series — official
- Xiaomi MiMo — MiMo-V2.6-Pro-RL model card — official
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.