Xiaomi MiMo-V2.6-Pro Tops Open-Source Models with 1M Context Omnimodal Release

MiMo-V2.6-Pro is a sparse mixture-of-experts omnimodal model from Xiaomi featuring 1.02 trillion total parameters and a 1 million token context window.

Xiaomi launched the MiMo-V2.6 series on September 22, 2026, with weights made available the previous day on September 21. The company positioned the release as a significant advancement in open-source artificial intelligence by providing models that rival closed-source systems in performance while remaining accessible under an MIT license. This move comes at a time when the industry sees increasing interest in open models that offer transparency and customizability. The dual release of Pro and Flash variants caters to different user needs, from high-performance computing to efficient inference on more modest hardware setups. The inclusion of a distilled 9B model in the release package further supports accessibility for a wider range of developers and researchers who may not have access to extensive computational resources. Availability through multiple platforms ensures that users can choose the deployment method that best suits their requirements, whether through direct downloads or API integrations.

What new elements does the MiMo-V2.6 release bring to the frontier models landscape?

The MiMo-V2.6-Pro and its Flash counterpart introduce several advancements over prior iterations in the series. Both models handle inputs across text, image, video, and audio modalities while generating text outputs. This omnimodal design allows for integrated processing of diverse data types in a single model instance without requiring separate specialized pipelines. A standout feature is the one million token context window, which enables handling of extensive documents, long conversations, or complex multimodal sequences without truncation. The models also support up to 128,000 output tokens. These specifications position the series competitively against proprietary offerings that often limit context lengths or charge premium rates for extended use. The open release of weights under the MIT license further distinguishes the effort by removing barriers that typically restrict modification and redistribution in closed ecosystems.

The accompanying technical report details evaluation results across standard benchmarks and provides the full set of over 7,000 RL task environments used during development. This level of disclosure allows independent researchers to replicate portions of the training pipeline or extend the work into new domains. Distribution on Hugging Face alongside OpenRouter integration broadens immediate access for both individual developers and enterprise teams. The combination of scale, modality support, and licensing terms creates a foundation for applications in areas such as long-form content analysis and multimodal agent development that previously required proprietary access.

How do MiMo-V2.6-Pro and MiMo-V2.6-Flash differ in architecture and scale?

MiMo-V2.6-Pro employs a sparse mixture-of-experts architecture with 1.02 trillion total parameters, of which 42 billion are active during inference. This design balances high capacity with computational efficiency by activating only relevant experts for each token. The Flash variant scales down to 309 billion total parameters and 15 billion active ones, targeting scenarios where speed and lower resource demands take priority over maximum capacity. Both variants maintain the same 1 million token context and 128,000 token output limits as well as identical modality support. The architectural choice reflects a deliberate strategy to serve distinct segments of the developer community with a single release cycle.

Specifications of the MiMo-V2.6 model variants released by Xiaomi

Variant	Total Parameters	Active Parameters	Context Window	Output Tokens
MiMo-V2.6-Pro	1.02 trillion	42 billion	1 million	128,000
MiMo-V2.6-Flash	309 billion	15 billion	1 million	128,000

What benchmark results support the performance claims for these models?

On the Artificial Analysis Intelligence Index version 4.3, MiMo-V2.6-Pro achieved a score of 46.32. This places it at the top among open-source models, exceeding the scores of Kimi K3 and Qwen3.8 Max according to the release materials. The score reflects comprehensive evaluations across multiple intelligence metrics including reasoning, knowledge retrieval, and multimodal understanding. Availability of the models on platforms like Hugging Face facilitates independent verification and further fine-tuning by the community. The result demonstrates that open-source efforts can close the gap with leading closed models when substantial compute is applied transparently.

The benchmark leadership is accompanied by detailed evaluation tables in the technical report that break down performance by category. These granular results enable users to assess suitability for specific use cases such as long-context reasoning or cross-modal tasks. The open weights allow direct comparison against other models hosted on the same platform, increasing confidence in the reported numbers through community scrutiny.

How was reinforcement learning scaled during the development of MiMo-V2.6?

The Xiaomi MiMo Team completed reinforcement learning training consisting of 30 steps in under six days. This process utilized approximately 750,000 trajectories. The effort incurred costs of about 2.62 million dollars for the Pro model and 0.85 million dollars for the Flash version. The relatively low cost compared with typical closed-model RL runs highlights the efficiency of the chosen task environments and distributed training setup. Public release of the training framework and task environments invites external contributions that could further reduce future iteration expenses.

1. Preparation of over 7,000 RL task environments for verifiable complex tasks
2. Execution of training across distributed compute resources to achieve 30 steps rapidly
3. Integration of exploration and feedback mechanisms to expand capability frontiers
4. Release of the full training framework alongside the models for community use

The RL approach focuses on verifiable outcomes to guide capability growth through iterative feedback. By making the environments and framework public, Xiaomi enables others to continue scaling from the released checkpoints. This transparency contrasts with many proprietary training runs that remain internal and undocumented beyond high-level summaries.

What are the pricing structures and distribution channels for MiMo-V2.6?

API pricing remains consistent with the V2.5 generation. The Pro model charges 0.435 dollars per million tokens for cache miss input and 0.87 dollars for output. The Flash variant offers lower rates at 0.14 dollars input and 0.28 dollars output per million tokens. These rates position the models as cost-effective options for high-volume applications while maintaining profitability for the provider. The unchanged pricing from the prior version signals stability in the commercial offering.

Models and associated resources appear on Hugging Face under the XiaomiMiMo organization. Additional access comes through Xiaomi platforms and OpenRouter. The MIT license permits broad commercial and research applications without restrictive terms. Users can therefore deploy the models in production environments or modify them for specialized needs with minimal legal friction.

What market and stakeholder implications arise from this open-source release?

The release challenges the dominance of closed-source frontier models by providing high-performance alternatives at reduced costs. Developers gain the ability to run or fine-tune large-scale models locally or via affordable APIs, potentially accelerating innovation in agentic systems and retrieval-augmented generation applications. Startups that previously relied on expensive proprietary APIs can now experiment with equivalent capabilities under an open license, lowering barriers to entry for new products.

Stakeholders including startups and academic institutions benefit from the transparency of weights and training details. This could shift competitive dynamics as organizations evaluate open options against subscription-based services from other providers. The availability of extensive RL task environments may also spur collaborative research projects that build upon the released foundation rather than starting from scratch.

What expert perspectives have emerged regarding the MiMo-V2.6 series?

The announcement emphasizes scaling RL compute on verifiable tasks to allow continuous capability expansion through exploration and feedback loops.

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

What developments can be anticipated in the coming months for Xiaomi's AI efforts?

Further iterations may build on the released framework and task environments. The provision of extensive RL resources suggests ongoing community contributions could refine performance in specialized domains such as domain-specific reasoning or multimodal alignment. Integration with tools like MiMo Desktop may expand practical applications for end users beyond API access.

Observers will monitor adoption rates on OpenRouter and Hugging Face to gauge real-world impact. Continued releases of distilled variants and additional training artifacts could maintain momentum in the open-source segment. The overall trajectory points toward sustained investment in public RL scaling as a core strategy for capability advancement.