Xiaomi MiMo-V2.6-Pro Tops Open-Source Benchmarks with 1.02T MoE Release

MiMo-V2.6-Pro is a 1.02-trillion-parameter sparse MoE model with 42 billion active parameters per token developed by Xiaomi and open-sourced under the MIT license.

Xiaomi announced the release of MiMo-V2.6-Pro and MiMo-V2.6-Flash on September 22, 2026, making the models available with open weights under the MIT license. The release follows a public RL training dashboard that allowed observers to track the process in real time. This approach to transparency in training marks a departure from typical closed development cycles in frontier model releases. The models target high performance in reasoning and agent tasks while maintaining low training costs relative to closed alternatives. By open sourcing the weights along with the full technical report, training environments, and RL code, Xiaomi enables broader community participation in further development.

What background led to the MiMo-V2.6 series release?

Xiaomi has pursued an exploration of reinforcement learning scaling as a path to self-improvement in large models. The company focused on verifiable complex tasks to allow the model to expand its capability frontier through exploration and feedback loops. Prior open-source models from other providers had established competitive benchmarks, prompting Xiaomi to emphasize cost efficiency during the RL phase. The public streaming of the training run provided visibility into a process completed in under six days using approximately 750,000 trajectories. This method contrasts with longer and more expensive pretraining runs common in the industry.

The strategy aligns with efforts to make frontier capabilities more accessible by reducing the compute barriers that often limit participation to well-funded organizations. Training costs reached approximately $0.85 million for the Flash variant and $2.62 million for the Pro variant according to the official announcement. Such figures illustrate how targeted RL can achieve strong results without the full expense of training from scratch. The open release of resources around the models further supports replication and extension by independent researchers.

What are the technical specifications of MiMo-V2.6-Pro and Flash?

MiMo-V2.6-Pro uses a sparse MoE architecture with 1.02 trillion total parameters and 42 billion active parameters per token. The Flash variant scales down to 309 billion total parameters with 15 billion active per token. Both models include a 1 million token context window that supports long-form reasoning and agent workflows. Native omnimodal input handling covers text, image, video, and audio within the same model weights. The design prioritizes inference efficiency through selective activation of experts during forward passes.

Comparison of key specifications for the MiMo-V2.6 series models

Model	Total Parameters	Active Parameters	Context Length	Modalities	License
MiMo-V2.6-Pro	1.02 trillion	42 billion	1 million tokens	Text, Image, Video, Audio	MIT
MiMo-V2.6-Flash	309 billion	15 billion	1 million tokens	Text, Image, Video, Audio	MIT

The sparse activation pattern reduces the effective compute required at inference time compared to dense models of similar total size. This architectural choice supports deployment on a wider range of hardware while preserving performance on complex tasks. The models were trained with a focus on scaling RL compute rather than solely increasing parameter count. Documentation on Hugging Face confirms the exact parameter counts and context capabilities for each variant.

How does performance compare on intelligence benchmarks?

The score places the Pro model at the top of open-source entries on the index while remaining competitive with select closed models. Performance gains trace directly to the RL phase that emphasized verifiable tasks and feedback mechanisms. The index aggregates multiple capability evaluations into a single comparable metric. Xiaomi stated that the result establishes MiMo-V2.6-Pro as the strongest open-source model to date on this particular benchmark.

The Flash variant offers a lighter option for scenarios where lower active parameter counts suffice. Both models demonstrate that efficient RL can close gaps with larger or more expensive systems. Community evaluation on platforms such as Hugging Face will provide additional data points beyond the initial index score.

What market and stakeholder implications follow the release?

The MIT license removes financial and legal barriers that previously restricted access to similar capability levels. Enterprises can now experiment with 1 million token context and omnimodal features without incurring per-token fees or usage restrictions. Lower training costs of $2.62 million for the larger model suggest a replicable template for organizations seeking to develop custom reasoning systems. The open release of RL code and environments accelerates experimentation across academic and commercial labs.

Stakeholders in agent development gain immediate access to models optimized for complex task chains. Competition among open-source providers may intensify as performance benchmarks continue to rise. The public training dashboard sets an expectation for greater transparency in future frontier releases from multiple organizations.

What reactions have experts expressed about the announcement?

Today, we are releasing and open-sourcing the MiMo-V2.6 series. This marks a key step in our exploration of the RSI path: scaling RL compute on verifiable, complex tasks, so the model can continuously expand its capability frontier through exploration and feedback.Xiaomi MiMo Team

The statement highlights the strategic emphasis on continuous capability expansion through reinforcement learning. Observers note that the accompanying open resources lower the threshold for independent verification of results. The combination of benchmark leadership and cost efficiency draws attention from both researchers and product teams focused on agent applications.

What developments are anticipated next for the MiMo series?

Subsequent releases may apply the same RL scaling approach to additional task domains and larger trajectory counts. Community contributions enabled by the open code and environments could produce specialized fine-tunes or architectural variants. Xiaomi may publish further details on the training environments to support broader replication efforts.

Continued public dashboards could become a standard practice for documenting RL runs in open-source projects. The models provide a foundation for testing new agent frameworks that leverage the long context and multimodal capabilities. Ongoing evaluations will determine how the performance edge holds against evolving closed and open competitors.

1. Release of full technical report and weights on Hugging Face
2. Open sourcing of training environments and RL code
3. Public streaming of the reinforcement learning run
4. Provision of resources to support community replication and extension