Xiaomi just released and open-sourced its MiMo-V2.6 series, with models positioned to compete with leading closed-source offerings on agentic tasks. But the more significant release is the methodology. The accompanying technical report details a scaled-up reinforcement learning (RL) strategy, making a public case that the path forward is not just bigger base models, but better, continuous improvement through RL at scale.
what changed: scaling rl in public
The core idea behind the MiMo-V2.6 series is scaling three components of reinforcement learning in tandem: RL compute, the diversity of task environments, and grader compute. Instead of discrete training runs for different domains, the approach uses one mixed RL run across coding, general agent tasks, visual tasks, and cybersecurity. This produced two main omni-modal models: MiMo-V2.6-Pro, the flagship, and MiMo-V2.6-Flash, a smaller model optimized for cost and speed.
The Pro model, a sparse mixture-of-experts architecture with over a trillion total parameters, now stands as one of the most powerful open-source models available, particularly on benchmarks that measure long-horizon engineering tasks. The company has fully open-sourced the weights and technical report for both the Pro and Flash models. They also released over 7,000 high-quality RL task environments covering software engineering, web design, and knowledge work, providing a valuable resource for others building agentic systems.
the engineering details that matter
The technical report provides specific details about the training process. The RL phase consumed massive compute, processing 1,568 prompts per training step, with each prompt generating 16 attempts. This high-throughput, asynchronous training pipeline allowed the model to learn from billions of tokens at each step within a context length of up to one million tokens.
A key improvement was in the quality of the feedback signal. The team moved beyond simple correctness verifiers to a more sophisticated "groupwise agentic grading" system. This grader analyzes execution traces to provide more accurate reward signals for long-running tasks, steering the model toward more token-efficient solutions. This investment in grader compute represented a significant portion of the total RL cost but was critical for keeping the training process stable and preventing reward hacking.
For builders who want to run the model directly, the Hugging Face page provides configuration details. Serving the Pro model can be done using a containerized environment with a library like vLLM.
# Example command to serve the model using a pre-built container
docker pull vllm/vllm-openai:mimov25-cu129
vllm serve XiaomiMiMo/MiMo-V2.6-Pro-RL \
--tensor-parallel-size 8 \
--trust-remote-code \
--gpu-memory-utilization 0.95 \
--max-model-len auto \
--reasoning-parser mimo \
--tool-call-parser mimo \
--enable-auto-tool-choice \
--generation-config vllm
This configuration highlights the model's native support for features like automatic tool choice, which is essential for agentic workflows.
the so-what for builders
The release of MiMo-V2.6 pushes the performance of open-source models further, particularly for complex, multi-step tasks that require an agent to interact with tools and environments. On some benchmarks for long-horizon software engineering, the Pro model's performance improved significantly over the course of its RL training.
While the closed-source labs are currently focused on a price war for high-volume inference, this release signals a different competitive front for open-source: the training methodology itself. By open-sourcing not just the model but also the RL environments, the release provides a toolkit for teams building their own specialized agents. It allows builders to replicate, fine-tune, and extend the RL process on their own data and tasks. This is less about providing a drop-in replacement for a commercial API and more about providing the foundation for building custom, self-improving systems.
Top comments (0)