Xiaomi’s September 22 MiMo announcement centers on a larger package that the company is making available to researchers and developers. The September 22 release includes model weights, a technical report, reinforcement-learning materials, and access paths for the MiMo-V2.6 series. That moves the story from how Xiaomi trained the models to what outside users can examine and build with.
Xiaomi’s official MiMo release notice says the company released and open-sourced MiMo-V2.6-Pro and MiMo-V2.6-Flash, along with MiMo-V2.6-Distill-Qwen-9B and related reinforcement-learning research resources. TechNode reported the same release and summarized Xiaomi’s stated performance and training figures.
The word open-sourced should be used with care. Xiaomi says it has released the weights, technical report, training environments, and RL code, but this article does not independently assess licensing terms. The practical news is that a broader set of artifacts is now available, not that every part of Xiaomi’s training and deployment stack has become public.
The release package is the main event
The series has two native multimodal models, MiMo-V2.6-Pro and MiMo-V2.6-Flash. Xiaomi also says it has released MiMo-V2.6-Distill-Qwen-9B, a smaller model in the package, and a collection of reinforcement-learning resources. The company directs users to its public collection at https://huggingface.co/collections/XiaomiMiMo/mimo-v26 and describes a release that includes the technical report, training environments, and RL code.
That level of disclosure gives developers more than an API endpoint. It creates a chance to study how Xiaomi frames the relationship between model design, task environments, and post-training. Xiaomi says the release includes more than 7,000 high-quality RL task environments across software engineering, vulnerability reproduction, knowledge work, and web design and development. That count is Xiaomi’s own description of the released resources.
EastFrontier’s September 19 report, Xiaomi Opens a Window Into MiMo V2.6 RL Training, examined the training-process disclosure that preceded this announcement. The new development is the release layer: Xiaomi is now giving outside users access to a model family and supporting research materials rather than only describing the process behind them.
Company-reported scores need context
Xiaomi places MiMo-V2.6-Pro at 46 points on the Artificial Analysis Intelligence Index. TechNode’s summary gives Kimi K3 a 44-point comparison and GLM-5.3 a 45-point comparison, while Xiaomi’s own release uses a slightly different comparison set. The safest reading is that Xiaomi is using those figures to position the Pro model, not that every comparison has been independently replicated for this story.
The company says Pro and Flash each underwent 30 reinforcement-learning steps in fewer than six days, drawing on roughly 750,000 trajectories. Xiaomi gives approximate training costs of US$2.62 million for Pro and US$850,000 for Flash. Those details are useful release context, but they should remain Xiaomi-reported metrics rather than the center of the article.
Why? A model’s stated cost and benchmark score do not resolve its real-world value. Users still need to inspect what was trained, what is public, how the model behaves on their workloads, and whether the tools and environments transfer beyond a curated demonstration. Xiaomi’s disclosure makes that scrutiny more possible, but it does not replace it.
The same restraint applies to the company’s capability examples. Xiaomi says the release supports 3D spatial reasoning, multimodal perception, and computer-use tasks, with examples including game or world construction, Blender modeling, robotic-arm simulation, and graphical-interface work. These are described capabilities and examples, not universal guarantees of autonomous performance.
An expanding MiMo line, with a clearer division of roles
MiMo-V2.6-Pro and MiMo-V2.6-Flash are the central released models. MiMo-V2.6-Pro-Ultraspeed is different: Xiaomi describes it as a faster service mode with output speed advertised at up to 20 times, rather than one of the two core open-weight releases. The company also says the MiMo-V2.6 models are available through its API.
The release separates Xiaomi’s stated open research-facing package from its API access and MiMo-V2.6-Pro-Ultraspeed service mode. Xiaomi describes Ultraspeed as a faster service option, with output speed advertised at up to 20 times. That claimed speed should remain attributed to Xiaomi and should not be confused with the public model-weight release. The company also says the MiMo-V2.6 models are available through the Xiaomi MiMo API, a separate access route from the listed research materials.
EastFrontier’s September 19 report, Xiaomi Opens a Window Into MiMo V2.6 RL Training, covered the earlier training-process disclosure. This release adds model weights, a technical report, training environments, and RL code to the materials Xiaomi says it is making available. The release does not establish that every component of Xiaomi’s training or deployment stack is public. Xiaomi’s notice says the release includes a technical report, training environments, and RL code; it also names MiMo-V2.6-Distill-Qwen-9B among the materials. Those are the bounded facts that distinguish the September 22 release from a conventional model-performance announcement.
The near-term test will be external use. Researchers will look at the weights, code, environments, and technical documentation. Developers will decide whether the models fit their applications. Xiaomi has supplied a public package and a set of company-reported performance claims. The next stage belongs to the people who can put both to work.