Public real-time access to the reinforcement learning process for the MiMo V2.6 series now includes a dashboard showing internal metrics and daily expenditures. The Pro model, featuring 1.02T total parameters and 42B active parameters, costs $493,000 per day to train, while the 309B parameter Flash model costs $247,000 per day. This compute run processes approximately 2B tokens per step with a 25,000 batch size, utilizing 1,568 prompts and 16 rollouts in an asynchronous setup.

The training system utilizes multi-task agentic RL mixed across multiple harnesses and agentic in-group credit assignment with rubric-based rewards. MiMo will open-source the implementation details piece by piece over the coming weeks. The public dashboard provides telemetry on batch data and harness composition, having already recorded more than $1 million in total expenses for the current run.