Xiaomi releases MiMo-V2.6-Pro, an open omnimodal model with RL-driven training and cost-efficient frontier performance

Xiaomi, traditionally known as the 'Apple of China' for its consumer electronics, has entered the frontier AI model space with the release of MiMo-V2.6-Pro, an open-weights omnimodal model. The company describes it as its most capable model to date, positioning it as a top performer on the Artificial Analysis Intelligence Index (46).

The model’s architecture includes 1.02T total parameters and 42B active parameters, with cost efficiency highlighted at $0.13 per Intelligence Index task. Independent analysis reported token-level costs of $0.435 per million input tokens and $0.87 per million output tokens, underscoring its competitive pricing relative to closed alternatives.

Training for MiMo-V2.6-Pro reportedly cost $3M, a figure cited in the announcement. The model’s RL training stack scaled compute across three axes: larger batches (up to 1,568 samples per update), longer contexts (up to 1 million tokens), and multi-task environments spanning coding, general agents, visual and cyber tasks, and music generation. The RL pipeline also incorporated richer environments and more grader compute to refine reward signals and reduce token usage per task.

Xiaomi open-sourced key components of the training stack, including the RL environments, training recipes, and technical report. However, the full set of 7,000+ task datasets was not released. The company also introduced two companion models: MiMo-V2.6-Flash, optimized for balance between intelligence, efficiency, and cost, and MiMo-V2.6-Pro-UltraSpeed, which claims up to 20x faster output at equivalent quality.

The release reflects a broader trend in China’s AI ecosystem, where labs are rapidly iterating on open-weight models. Xiaomi’s approach emphasizes transparency in RL training, with former DeepSeek engineer Fuli Luo publishing final RL training runs live and detailing internal metrics. The technical report and environment code are available under the MIT license, further lowering barriers to adoption.

Independent observers noted that Xiaomi’s RL scaling—reportedly requiring 130 hours, 75 billion tokens, and $2.6M for the RL phase alone—demonstrates that post-training can be a cost-effective route to frontier-adjacent performance. The training leveraged JAX and TPU infrastructure, with scaling described as a configuration change rather than a code rewrite, suggesting efficiency gains in compute utilization.