Xiaomi MiMo-V2.5 Open Source Models Put New Pressure On Agentic AI Costs

Advertisement

Xiaomi has released MiMo-V2.5 and MiMo-V2.5-Pro as open source AI models, widening its push beyond smartphones, electric vehicles, and connected devices into the fast moving market for agentic AI systems.

The two models were released under the MIT License, a permissive software license that allows developers and companies to use, modify, fine tune, and deploy the models commercially. That detail matters because many enterprise teams are now weighing not only model quality, but also control, cost, and deployment flexibility.

MiMo-V2.5 is positioned as the broader multimodal model, while MiMo-V2.5-Pro is aimed more directly at long running agentic tasks, software engineering workflows, and systems that require many tool calls over time. According to the provided benchmark details, the Pro version reached a 63.8 percent success rate on ClawEval while using about 70,000 tokens per trajectory.

That efficiency is the main point of the release. Agentic AI systems do not simply answer one question and stop. They may read files, write code, revise outputs, manage email, schedule tasks, or operate inside third party tools. Those workflows can become expensive when every step consumes tokens.

Xiaomi’s claim is that MiMo-V2.5-Pro can complete those extended tasks while using fewer tokens than several closed frontier models named in the provided material. For developers, that can change the economics of building persistent agents, especially when usage based billing is becoming more common across AI services.

The architecture also reflects that focus. MiMo-V2.5 uses a sparse Mixture of Experts design with 310 billion total parameters and 15 billion active during inference. MiMo-V2.5-Pro is described as a larger 1.02 trillion parameter Mixture of Experts model with 42 billion active parameters.

The Pro model’s training appears more narrowly aimed at sustained action. The supplied information says it is built for long horizon coherence, complex software engineering, and better memory management inside agent frameworks. Xiaomi’s reported examples include a compiler project in Rust, a desktop video editor, and an analog engineering optimization task.

The release also carries a clear business signal. Xiaomi is not presenting MiMo only as a research model. The company is making it available through Hugging Face, API access, and subscription style token plans. It has also offered free cache writing for a limited period and a large free token grant for builders.

For companies outside China, the open weights may be the more important route. The provided material notes that U.S. enterprises may face compliance or regulatory concerns when relying directly on Chinese technology services. Running the model locally or inside a private cloud could give those users more control while still taking advantage of the open source release.

The timing is important. AI software costs are moving away from simple monthly subscriptions and toward token based consumption. In that environment, models that can run privately and handle long agentic sessions at lower cost become more attractive.

Xiaomi’s move does not settle the race between open and closed AI systems. It does, however, sharpen the question facing developers and enterprise buyers. If an open model can deliver strong agentic performance with flexible licensing and lower operating costs, the value of proprietary systems will have to be proven more clearly in real production work.

For now, MiMo-V2.5 and MiMo-V2.5-Pro give Xiaomi a serious place in the open source AI conversation. The release shows that the next phase of competition may not be only about bigger benchmarks, but about which models can do useful work reliably, affordably, and under terms that developers can actually live with.