Xiaomi Open-Sources MiMo-V2.5 AI Models with 100 Trillion Token Developer Programme

A Major Push into Open-Source AI

Xiaomi has released the full weights of its MiMo-V2.5 model family under the MIT Licence, making two of its most capable AI models available for commercial use, fine-tuning, and local deployment. The release follows a public beta that began on 23 April and represents a significant move by the Chinese hardware giant into the competitive open-source AI landscape.

Two Models, One Mission

The family includes MiMo-V2.5, a 310-billion-parameter native omnimodal model capable of processing text, images, video, and audio, and MiMo-V2.5-Pro, a 1.02-trillion-parameter Mixture-of-Experts model with 42 billion active parameters. Both support context windows of up to one million tokens.

MiMo-V2.5-Pro is Xiaomi's flagship, built for complex agentic workflows and software engineering tasks. On the ClawEval benchmark, it achieved a 63.8% success rate whilst consuming roughly 40-60% fewer tokens per trajectory than comparable models from Anthropic, Google, and OpenAI. On SWE-bench Pro, the model scored 57.2, placing it alongside Claude Opus 4.6 and GPT-5.4. Xiaomi demonstrated its long-horizon capabilities by having it autonomously build a complete SysY compiler in Rust, passing all 233 tests after 672 tool calls in 4.3 hours.

AMD announced day-zero support for MiMo-V2.5-Pro on its Instinct GPUs, signalling immediate infrastructure readiness for deployment.

MiMo Orbit: 100 Trillion Tokens for Developers

Alongside the open-source release, Xiaomi launched MiMo Orbit, an initiative to seed its developer ecosystem. The programme's centrepiece is a 100-trillion-token grant distributed free to approved developers over 30 days beginning 27 April, with individual applicants eligible for up to 1.6 billion credits valued at roughly 659 CNY, or approximately $91. A second component, the Agent Ecosystem Co-building initiative, offers free model integration services to agent framework teams worldwide.

Competing on Cost

The token efficiency claims carry practical weight for production deployments. MiMo-V2.5-Pro is priced at $1.00 per million input tokens and $3.00 per million output tokens through Xiaomi's API, a fraction of the cost of Claude Opus 4.6 at $5 and $25 respectively, or GPT-5.5 at $5 and $30. Combined with lower token consumption per task, Xiaomi is positioning MiMo as a cost-competitive alternative for developers building agent pipelines at scale.