Xiaomi Releases MiMo-V2.6 Models as Its Pro Version Tops an Open-Weight Ranking

Xiaomi has released MiMo-V2.6, a new open-weight AI model family led by MiMo-V2.6-Pro, which VentureBeat reports is now the highest-scoring open-weight model on Artificial Analysis’ Intelligence Index. The launch combines a trillion-parameter flagship, a lower-cost Flash model, and a speed-focused serving version—giving developers more than one route into Xiaomi’s latest models.

MiMo-V2.6-Pro is available under the MIT license and can be downloaded through Hugging Face, according to VentureBeat. Xiaomi also offers the model through its API at $0.435 per million uncached input tokens and $0.87 per million output tokens. That puts a model positioned near the top of a third-party ranking behind both a downloadable release and a paid hosted service.

A flagship built for unusually large inputs

The Pro model is a sparse mixture-of-experts system: it has 1.02 trillion parameters in total, but Xiaomi says 42 billion are active during inference. It accepts text, images, audio, and video, with a reported context window of one million tokens—the amount of material it can take into a request at once.

Artificial Analysis gave Pro an Intelligence Index score of 46, tying the reported score for xAI’s Grok 4.7 and exceeding the scores listed for Grok 4.6, Gemini 3.8 Flash, and two DeepSeek V4.1 models. The index result is a competitive marker, not a demonstration that the model wins every kind of task.

Xiaomi’s response is a split product line

Rather than ask every customer to use the flagship, Xiaomi released MiMo-V2.6-Flash alongside it. Flash has 310 billion total parameters and 15 billion active parameters, according to Xiaomi’s reported specifications. Its lower pricing targets high-volume work while retaining the family’s advertised long-context and multimodal features.

The third release, MiMo-V2.6-Pro-UltraSpeed, is the company’s answer for workloads where output rate is central. Xiaomi says it can generate at up to 20 times the normal Pro output speed. That is a company performance claim, and the evidence supplied here does not provide comparable latency or throughput tests across deployment conditions.

The training bill shows what Xiaomi optimized for

Xiaomi says both Pro and Flash received 30 reinforcement-learning steps covering roughly 750,000 trajectories in under six days. It reports spending about $2.62 million on Pro’s post-training run and about $850,000 on Flash’s. Reinforcement learning here means improving a model by scoring its attempts at tasks, including longer sequences of actions rather than only short answers.

The company also says it tried to limit reward hacking, where an agent finds a way to satisfy a grader without doing the intended work. Xiaomi reported confirmed reward-hacking trajectories below 2% for both models in the final run. That figure describes Xiaomi’s own training measurement, not an independent safety assessment.

The questions a ranking cannot settle

• Whether the million-token window reliably retrieves and reasons over information throughout a long prompt.
• Whether tool-using agents recover cleanly from failures instead of producing brittle or wasteful chains of actions.
• How Pro and Flash compare with closed rivals on evaluations where Xiaomi says leading proprietary models remain ahead.

Xiaomi is not claiming a sweep of every test. VentureBeat reports that Claude Opus 5 and GPT-5.6 Sol still lead on some evaluations cited by Xiaomi. The next move for prospective users is therefore straightforward: treat MiMo-V2.6’s price, license, and index position as reasons to evaluate it, rather than proof that it will be the best model for a production workflow.