ANALYSIS A consumer-electronics company just matched the benchmark ceiling that the most expensive proprietary labs occupy, and it did so for a reported training cost that rounds to a rounding error on their R&D budgets. Xiaomi's MiMo-V2.6-Pro scored 46 on the Artificial Analysis Intelligence Index, tying xAI's Grok 4.7 and topping every other open-weight model on the leaderboard5,13. The RL phase that produced that result cost $2.62 million and ran for less than six days12. For closed-model labs whose commercial moats rest on capability gaps, the math just got uncomfortable.
Why it matters
The release redraws two boundaries at once. First, the open-weight frontier now sits at the same Intelligence Index score as a leading proprietary system. Second, the cost to reach that score has dropped to a figure accessible to well-funded university labs, not just hyperscalers. Xiaomi reported the full RL run consumed roughly 750,000 trajectories across 30 steps, burning 3.5 to 3.7 billion tokens per step, for a total spend of $2.62 million on Pro and $850,000 on Flash3. Those figures, if reproducible by third parties, compress the timeline in which any motivated team can field a frontier-class model under an MIT license.
The distillation controversy adds a second dimension. On September 10, 2026, Anthropic published a report naming Xiaomi alongside DeepSeek, Moonshot AI, and Zhipu as firms it believes routed user queries through Claude to harvest training data9. Eleven days later, MiMo-V2.6-Pro topped the open-weight leaderboard. Xiaomi has not addressed the allegation publicly10, and no evidence has surfaced tying MiMo-V2.6 specifically to Claude-derived training data. The juxtaposition sharpens a question closed labs will have to answer with more than blog posts: if open-weight competitors can reach parity at a fraction of the cost, the distillation accusation functions as a policy argument for restricting access, not a technical one for maintaining superiority.
The big picture
MiMo-V2.6-Pro is a sparse mixture-of-experts model: 1.02 trillion total parameters, roughly 42 billion active11. It ships with a 1-million-token context window and native multimodal capabilities spanning text, images, audio, and video4. Xiaomi claims performance "on par with Opus 5 and GPT-5.6 Sol across most agent benchmarks"6. On DeepSWE v1.1, the vendor-reported score is 71.9%, trailing Claude Opus 5 at 74.0%. On the Artificial Analysis Intelligence Index v4.3.2, MiMo-V2.6-Pro sits at 46, ahead of GLM-5.3 at 45 and Kimi K3 at 44, and level with Grok 4.7 (xHigh). Leading closed-source models scored 53 on the same index.
The model is available on Hugging Face under an MIT license, alongside RL code and more than 7,000 task environments. That packaging matters as much as the weights themselves: it gives external teams the scaffolding to replicate or extend Xiaomi's RL recipe rather than merely run inference.
The team behind the release is led by Fuli Luo, formerly of DeepSeek1. Luo's group livestreamed the final RL training runs before the release, an unusual degree of transparency that drew public attention. The move positions Xiaomi not just as a model builder but as an open-research institution, a brand play that a hardware company can afford in ways a pure-play API vendor cannot.
Between the lines
Xiaomi's technical report details a pipeline built on fully asynchronous Group Relative Policy Optimization at 1,568 prompts with 16 rollouts per step. The RL training mixed coding, general agent, visual, and cybersecurity tasks. Reward-hacking defenses span reward design, adversarial evaluation, anomaly detection, and cross-checking between verifiers.
Xiaomi charges $0.435 per million input tokens and $0.87 per million output tokens for MiMo-V2.6-Pro. Artificial Analysis rates MiMo-V2.6-Pro as the cheapest model it tracks, at $0.13 per task. Anthropic cut Claude Opus 5.5 to $4 and $20 per million tokens on the same day. The architecture's deliberate bet on sparse activation, with only 42 billion parameters active out of 1.02 trillion total, is what makes that pricing possible: inference costs stay low even as the model's representational capacity scales.
The gap between that 46 and the top closed-source score of 53 on the same index is real, and Xiaomi's own benchmark tables acknowledge it. But the competitive pressure runs in one direction: closed labs must now justify premium pricing against an MIT-licensed alternative that matches or exceeds several proprietary peers.
What's next
Xiaomi's UltraSpeed variant, which the company says delivers up to 20x faster output at the same quality, is rolling out now2. Artificial Analysis measured 129.7 output tokens per second and a 2.17-second time to first token through Xiaomi's API. Xiaomi reported vendor-side DeepSWE v1.1 lifts from 48.8 to 65.7 for Flash and from 58.4 to 72.6 for Pro, though those gains remain unconfirmed by third parties. Independent reproduction will be the real test; Xiaomi has published the environments and code to make that possible. The weights are live, the license is MIT, and the cost figure is on the record. Anthropic's distillation allegation, still unaddressed by Xiaomi, ensures that the policy debate over open-weight proliferation will track this model's adoption closely.