Xiaomi released MiMo-V2-Pro, a 1-trillion parameter foundation model, on March 18, 2026, positioning it as a direct competitor to models from OpenAI and Anthropic at roughly one-sixth to one-seventh the API cost. The model is led by Fuli Luo, a veteran of the DeepSeek R1 project, who described the release as a “quiet ambush” on the global AI frontier.
Architecture Built for the Agent Era
MiMo-V2-Pro uses a sparse architecture that activates only 42 billion of its 1 trillion total parameters during any single forward pass. This design makes it approximately three times larger than its predecessor, MiMo-V2-Flash, while maintaining efficiency. The model features a 7:1 hybrid attention ratio, up from 5:1 in the Flash version, enabling it to manage a 1-million-token context window without the performance degradation common in large models.
A Multi-Token Prediction layer allows the model to generate multiple tokens simultaneously, reducing latency during complex reasoning tasks. Luo stated these architectural decisions were made months in advance to provide a structural advantage as the industry shifted toward agentic workflows.
MiMo-V2-Pro Benchmark Results and Third-Party Verification
On GDPval-AA, a benchmark measuring performance on real-world agentic tasks, MiMo-V2-Pro achieved an Elo score of 1426. This places it ahead of Chinese peers GLM-5 (1406) and Kimi K2.5 (1283), though it trails Claude Sonnet 4.6 (1633). The third-party organization Artificial Analysis ranked MiMo-V2-Pro at number 10 on its global Intelligence Index with a score of 49, placing it in the same tier as GPT-5.2 Codex and ahead of Grok 4.20 Beta.
Key metrics from Artificial Analysis show notable improvements over MiMo-V2-Flash, which scored 41 on the same index. The Pro model reduced its hallucination rate to 30%, down from 48% in the Flash version. It also required only 77 million output tokens to complete the full Intelligence Index, compared to 109 million for GLM-5 and 89 million for Kimi K2.5. On ClawEval, a benchmark for agentic scaffolds, MiMo-V2-Pro scored 61.5, approaching Claude Opus 4.6 at 66.3 and outpacing GPT-5.2 at 50.0. In Terminal-Bench 2.0, it achieved 86.7.
Pricing and Availability
Xiaomi priced MiMo-V2-Pro at $1 per million input tokens and $3 per million output tokens for contexts up to 256,000 tokens. For contexts between 256,000 and 1 million tokens, pricing rises to $2 per million input tokens and $6 per million output tokens. Cache reads are priced at $0.20 per million tokens for the lower tier and $0.40 for the higher tier, while cache writes are temporarily free.
By comparison, Artificial Analysis reported that running its full Intelligence Index cost $348 using MiMo-V2-Pro, versus $2,304 for GPT-5.2 and $2,486 for Claude Opus 4.6. The model is currently available only through Xiaomi’s first-party API and does not support image or multimodal input. Xiaomi has indicated a separate MiMo-V2-Omni model is in development for multimodal use cases.
Enterprise Considerations and Security Risks
For enterprise teams evaluating artificial intelligence infrastructure, the model’s 1-million-token context window supports retrieval-augmented generation architectures that can process entire codebases or documentation sets in a single prompt. Its optimization for OpenClaw and Claude Code makes it a candidate for multi-agent coordination and long-horizon planning tasks.
However, cybersecurity teams must account for the expanded attack surface that comes with agentic capabilities, including terminal access and file manipulation. The absence of public model weights, unlike the Flash version, also limits the depth of internal security audits available to organizations handling sensitive data. Xiaomi has not yet confirmed a timeline for open-sourcing the Pro variant, with Luo stating it will happen “when the models are stable enough to deserve it.”
Background: Xiaomi’s Path to Frontier AI
Beijing-based Xiaomi is the world’s third-largest smartphone manufacturer and entered the automotive sector in the early 2020s with electric vehicles including the SU7 sedan and the YU7 SUV. The company’s background in hardware and software integration informs MiMo-V2-Pro’s design as a reasoning engine for complex systems. Artificial Analysis currently ranks the model second in China and eighth globally on established intelligence indices.