When we think of Xiaomi, it is normal that its mobile phones come to mind or, at most, its foray into electric cars with models like the SU7. However, what we have seen now points to a much more ambitious move: the company also wants to compete in the artificial intelligence race. It has done so with the launch of MiMo-V2-Proa model that, according to the data shared by the company itself, seeks to position itself close to the most advanced systems on the market, but with a very different focus on costs. And that changes the conversation quite a bit.
What Xiaomi proposes. The company presents its model as the “brain” of systems capable of executing complete tasks, not just responding to specific requests, which in the sector is known as agent-oriented models. According to official information, we are looking at an architecture that exceeds one trillion total parameters, although it only activates 42 billion in each execution, and that can work with contexts of up to one million tokens. On paper, this allows you to maintain long, complex processes without fragmenting them, something designed for large tasks and more demanding workflows.
Performance against the greats. If we look at the data, Xiaomi does not present its model as the best on the market, but as one that can compete in certain scenarios. In the GDPval-AA benchmark, oriented to real agent-type tasks, it reaches an Elo of 1426, surpassing Chinese models such as GLM-5 (1412) and Kimi K2.5 (1309), although it falls short of proposals such as Claude Sonnet 4.6, which marks 1633. The external reading is provided by Artificial Analysis, which assigns it a score of 49 on its intelligence index, which places it in the group of most competitive models on the market. The key is in that closeness in some benchmarks, not in general leadership.
The key to the price. This is where Xiaomi’s proposal changes the board. According to data collected by Artificial Analysis, running your IQ with this model costs approximately $348, compared to $2,304 for GPT-5.2 or the 2,486 of Claude Opus 4.6. It is not exactly the same comparison as the price per API use, but on both levels Xiaomi appears clearly below several Western rivals. In its own API, the company sets prices of $1 per million tokens for entry and $3 for exit in the range up to 256K, a lower rate than models such as Claude Sonnet 4.6 and Claude Opus 4.6 at the same level of use.
Beyond chat. What Xiaomi is proposing with this model is not only to improve the quality of the responses, but to change the type of work it can do. The company insists on moving from conversation to action, with a system capable of using tools, interacting with environments and completing chained tasks. In this context, it presents it as a model optimized for agentic scenarios and links it to frameworks such as OpenClawin addition to mentioning collaborations with OpenCode, KiloCode, Blackbox and Cline. On paper, this reinforces the idea of an AI designed to execute workflows and not just answer questions.
behind the scenes. Xiaomi enters the race with a model that, according to available data, is close to the major benchmarks in some scenarios, although without generally surpassing them. Where there does seem to be a clear bet is on the price, and that is where it tries to differentiate itself. The question is whether this balance between cost and performance is maintained outside of benchmarks and in real environments. We will have to wait to know if what the data shows is also projected in the real world.
Images | Xiaomi
In Xataka | China has immediately understood the future of the technology industry: “one-person companies” powered by AI
GIPHY App Key not set. Please check settings