Core content: From Since the open source on April 29, 2026, the cumulative downloads of Xiaomi MiMo-V2.5 series large models on the Hugging Face platform have exceeded 2 million times, becoming the world's most popular agent-specific open source model. With MIT open source licenses, leading task efficiency and a highly competitive cost structure, MiMo is rapidly evolving from a technology product to an infrastructure-level choice for companies to build AI agents.
Open source is freedom: MIT license ignites developer enthusiasm
The primary reason why the MiMo-V2.5 series can achieve such rapid ecological growth in just one month is that it adopts the standard MIT open source license-one of the most relaxed and accepted open source protocols currently available by enterprises. Developers do not need authorization:
Integrate models directly in commercial products;
Conduct secondary training based on private data;
Publish derivative models or modified versions;
Deploy on local servers or private cloud environments.
This strategy is related to manyIn sharp contrast, the so-called open source model that "only opens weights but adds usage restrictions." As a CTO of a domestic autonomous driving startup said: "We dare not bet our core business on closed-source APIs, but MiMo can fully control it. This is true production-level availability."
Efficiency crushing: Completing the same task reduces Token consumption by 40 - 60%
Xiaomi released simultaneously ClawEval agent mission evaluation benchmark shows that MiMo-V2.5-Pro ranks first among open source models, with a mission success rate of 63.8%. More importantly, it only takes about 70,000 tokens to complete a complete task trajectory, while Anthropic Claude Opus 4.6, Google Gemini 3.1 Pro and OpenAI GPT-5.4 require 1.20,000 - 180,000 tokens to complete similar tasks, which is 40% to 60% more.
In GitHub Copilot fully shifted to token quota billing on May 1. This efficiency advantage directly translates into significant cost savings. For companies that use agents at high frequencies, switching to MiMo can reduce monthly AI spending by more than half.
Measured shock: 4.3 hours to write a full score compiler, 11.5 hours to develop a video editor
The capabilities of MiMo-V2.5-Pro are not only reflected in benchmark scores, but also verified in real complex tasks:
SysY compiler (Rust implementation):
model in 4.3 The lexical analysis, syntax parsing, semantic check and RISC-V backend generation were completed independently within hours. The tool was called 672 times in total. Finally, the full score of 233/233 was obtained on the hidden test set-this task usually takes several weeks for computer science students.
Desktop video editor:
time-consuming 11.5 In an hour, the tool was called 1,868 times and 8,192 lines of code were produced, supporting multi-track timeline, editing, transition and export functions.
Analog circuit optimization:
In Under the TSMC 180nm process, the linear regulation performance of the Flipped-Voltage-Follower (FVF-LDO) regulator is improved by 22 times through ngspice simulation cycles.
These cases demonstrate MiMo-V2.5-Pro's unique "harness awareness" capabilities-the model can proactively manage memory, plan tool call sequences, and maintain contextual consistency in long task chains.
Dual-mode strategy: Covering both multimodal and agent scenarios
Xiaomi usesThe "dual-track strategy" accurately targets different needs:
MiMo-V2.5 (all-round multimodal version):
310 billion sparse parameters (MoE), approximately 15 billion per activation
Native support for text, image, and audio understanding
Suitable for content generation, visual question and answer, and cross-modal retrieval
MiMo-V2.5-Pro (Agent Expert Edition):
1.02 Trillions of parameters MoE, activation parameters reach 42 billion
Specially designed for"More than a thousand tool calls" design
using 7:1 local-global attention ratio, efficiently focusing on key information
Both support 1 million tokens are long context and no additional "context tax".
Cost revolution: API pricing is only 1/5 of Claude Opus
Xiaomi provides highly competitive API pricing (overseas):
MiMo-V2.5-Pro: Input $1.00 /million tokens (≤256K), output $3.00
MiMo-V2.5: Input $0.40, output $2.00
Compared with mainstream models,The price of MiMo-V2.5-Pro is only 1/5 of Claude Opus 4.7 and 1/2.5 of GPT-5.4. In addition, Xiaomi offers input prices as low as $0.20 for cache hit requests and waives cache write fees for a limited period of time.
Companies can also choose The "Token Plan" subscription plan, with an annual payment of up to 19.2 billion tokens, is specially designed for high-intensity developers.
Ecological collaboration: chip + framework + cloud full-stack support
As a connection super 8.23 With an ecological giant of 100 million smart devices, Xiaomi is deeply integrating MiMo into its "people × car × home" strategy:
in-vehicle Agent: SU7/YU7 smart cars have tested integrating MiMo for voice interaction and task scheduling;
IoT control hub: Mijia equipment group achieves cross-device collaborative automation through MiMo;
Hardware adaptation: Brother Flat HeadMass production of the "Guang-MiMo" dedicated chip, increasing the inference energy efficiency ratio by 2.1 times;
Software Ecosystem: Obtained SGLang, vLLM "Day-0" support, compatible with AWS, AMD, Suiyuan Technology and other domestic and foreign hardware.
Developer benefits: 100 trillion free Tokens
To accelerate ecological construction, Xiaomi announced that it will provide global developers 100 trillion free tokens credit can be used to experience 1M long context and agent tasks without financial risk. Project leader Luo Fuli (former core member of DeepSeek) said on the social platform: "The value of the model does not lie in the ranking, but in solving practical problems. Now, start building with MiMo!"
Editor's note:millet The explosive growth of the MiMo-V2.5 series marks the first time that China technology companies have achieved global leadership in agent task efficiency and open source ecosystem. It is not only a technological breakthrough, but also a paradigm battle of "open source versus closed source","controllable versus black box", and "predictable cost versus usage anxiety". When companies begin to deploy AI employees on a large scale, MiMo may become the underlying operating system of this intelligent revolution.