Xiaomi released MiMo V2.5 Pro on April 22, 2026, and within days the model had tied for first place among open-weights AI on the Artificial Analysis Intelligence Index , with model weights still to come, meaning the developer community has not yet had full access to what already ranks ahead of DeepSeek V4 Pro.
The result is worth pausing on. MiMo V2.5 Pro scored 54 on the Artificial Analysis Intelligence Index, tying with Moonshot's Kimi K2.6 as the highest-ranked open-weights model currently available. Its composite intelligence score of 1578 places it ahead of DeepSeek V4 Pro at 1554, GLM-5.1 at 1535, and MiniMax-M2.7 at 1514. That ranking is based on evaluations across reasoning, mathematics, coding, and knowledge , the full suite of capabilities that matter for real-world deployment. For a model from a consumer electronics company that most Western developers had not treated as a serious AI competitor eighteen months ago, this is not a minor benchmark result. It is a statement of position.
MiMo V2.5 Pro is a Mixture-of-Experts architecture with 1 trillion total parameters, 42 billion of which are active during inference , a design that delivers frontier-tier capability at a fraction of the compute cost of a dense model of equivalent scale. It supports a 1 million token context window, handles agentic task chains with more than 1,000 tool calls, and scores 57.2% on SWE-bench Pro, a coding benchmark that has become one of the more demanding real-world proxies for software engineering capability. On HLE, a reasoning evaluation that proprietary frontier models dominate, MiMo V2.5 Pro scores 34%, up six percentage points from its predecessor. On IFBench, which measures instruction following, it scores 80%, up eleven points. The improvement over MiMo V2 Pro , which was itself released just five weeks earlier on March 18 , is substantial across every meaningful dimension.
Pricing is where Xiaomi is making a deliberate strategic choice. MiMo V2.5 Pro is available via Xiaomi's first-party API at $1.00 per million input tokens and $3.00 per million output tokens. Running the full Artificial Analysis Intelligence Index evaluation costs $462 for MiMo V2.5 Pro, compared to $948 for Kimi K2.6 and $544 for GLM-5.1. The model also uses roughly 40 to 60% fewer output tokens than Claude Opus 4 on comparable agentic tasks, according to benchmark data cited in analysis published by Artificial Analysis. For enterprise developers running high-volume workloads, that token efficiency is not an academic point. It is a direct line to infrastructure costs.
The Open-Weight Announcement Changes the Calculus
What makes the MiMo V2.5 Pro release more than a leaderboard story is what has not happened yet. Xiaomi has publicly committed to releasing the model weights as open-source, which would make it the strongest freely available model in the world at the time of release , ahead of any existing open-weight alternative by the Artificial Analysis scoring metric. The developer community on Reddit's LocalLLaMA forum has been tracking the release closely, with the phrase "Weights are coming" spreading as shorthand for anticipation of what would be, if the benchmark performance holds after weights drop, a significant moment for local AI deployment. Open weights mean fine-tuning, local inference, and integration into products without per-token API fees , exactly the properties that made DeepSeek's earlier open releases so disruptive to Western labs' pricing assumptions.
The broader pattern here is hard to ignore. As Bloomberg recently reported, Chinese AI labs have systematically narrowed the gap with Western frontier models over the past eighteen months, with DeepSeek, Qwen, and now Xiaomi's MiMo series benchmarking at or above proprietary models from Anthropic, Google, and OpenAI on specific evaluations , while pricing their API access at a fraction of the rate. MiMo V2.5 Pro is not yet the best AI model in the world by every measure; its hallucination accuracy on the AA-Omniscience Index remains a weak point, with low accuracy (23%) despite a low hallucination rate (25%), suggesting the model is cautious rather than confident when answering factual questions. Proprietary frontier models still lead there. But on reasoning, instruction following, agentic task completion, and cost per unit of intelligence, Xiaomi is now operating at the top of the market. When the weights land, that assessment will be tested by every developer with a GPU.
Also read: Sam Altman says he is deeply sorry OpenAI did not flag Tumbler Ridge shooter to police • Intel's CPU business just proved that agentic AI has a second hardware winner and it isn't Nvidia • Meta and Microsoft just cut 20,000 jobs while spending hundreds of billions on AI. That is not a contradiction.
