- Xiaomi released MiMo-V2-Pro on March 18, 2026 — a 1 trillion parameter mixture-of-experts model with 42 billion active parameters and a 1 million token context window.
- The model scores 78.0% on SWE-bench Verified and 61.5 on ClawEval, placing it in the top tier globally for coding and agentic tasks.
- Xiaomi deployed the model anonymously as “Hunter Alpha” on OpenRouter, where users and reviewers mistook it for DeepSeek V4 before the company revealed its origin.
- Pricing sits at $1.00 per million input tokens and $3.00 per million output tokens, undercutting most competing frontier models.
What Happened
Xiaomi, the Chinese smartphone and electronics manufacturer, released MiMo-V2-Pro on March 18, 2026. The model is a 1 trillion parameter mixture-of-experts architecture with 42 billion active parameters per token — roughly three times larger than its predecessor, MiMo-V2-Flash. It supports a context window of up to 1 million tokens and includes a multi-token prediction layer for faster generation speed.
Before the official announcement, Xiaomi deployed the model anonymously as “Hunter Alpha” on the OpenRouter platform. Multiple users and reviewers mistook it for DeepSeek V4, one of the most anticipated unreleased models at the time. The confusion validated that a smartphone manufacturer could produce AI models competitive with the output of dedicated AI research laboratories.
Why It Matters
MiMo-V2-Pro demonstrates that powerful AI models are no longer exclusive to companies like OpenAI, Anthropic, or Google DeepMind. Xiaomi is primarily a hardware company — smartphones, IoT devices, smart home appliances, and electric vehicles. Building a proprietary AI model gives Xiaomi an intelligence layer that connects its entire product ecosystem without depending on third-party API providers or their pricing changes.
The model can run inference across Xiaomi’s entire product line: answering voice queries on smartphones, managing smart home systems, and powering in-car assistants for Xiaomi’s electric vehicles. Training the model is a large but fixed cost. Running it across hundreds of millions of devices at marginal cost after training fundamentally changes the economics of device intelligence for Xiaomi’s ecosystem.
Technical Details
MiMo-V2-Pro uses a hybrid attention mechanism with a 7:1 hybrid ratio, increased from the 5:1 ratio in earlier versions. The architecture combines mixture-of-experts routing with the hybrid attention design to balance computational efficiency against model capability.
On coding benchmarks, the model scores 78.0% on SWE-bench Verified, 71.7% on SWE-bench Multilingual, and 57.1% on Terminal-Bench 2.0. On general agent benchmarks, it scores 81.0 on PinchBench (ranking third globally), 61.5 on ClawEval (also third globally, approaching Claude Opus 4.6’s performance), and achieves 1426 Elo on GDPVal-AA.
For tool use and search tasks, the model reaches 96.8% on the telecom-focused tau-2 benchmark and 86.7% on DeepSearch QA-F1. Developer feedback during the anonymous Hunter Alpha testing period indicated performance that surpassed Claude Sonnet 4.6 in most coding scenarios. The model ranks eighth globally and second among Chinese large language models on the Artificial Analysis Intelligence Index.
Who’s Affected
Developers and enterprises looking for cost-effective AI models gain a competitive new option. Pricing is set at $1.00 to $2.00 per million input tokens and $3.00 to $6.00 per million output tokens, depending on context window size. Cache reads cost $0.20 to $0.40 per million tokens, with temporary cache writes offered at no charge. These rates undercut most frontier models from established AI labs by a significant margin.
Xiaomi joins Huawei, Samsung, and Apple as hardware companies that are now also AI model companies. This trend blurs the line between device manufacturers and AI infrastructure providers. For the dedicated AI labs, the emergence of competitive models from hardware companies increases pricing pressure and reduces the defensibility of proprietary model development as a standalone business.
What’s Next
MiMo-V2-Pro specializes in agentic workflows, coding tasks, frontend development, and complex multi-step reasoning with strong tool-calling accuracy. Xiaomi’s next step will likely be deeper integration of the model across its hardware ecosystem, particularly in its growing electric vehicle division where on-device AI enables voice assistants and autonomous driving features.
The primary limitation is ecosystem maturity. While the model’s benchmark scores are competitive with frontier offerings, Xiaomi lacks the developer tooling, documentation, API ecosystem, and enterprise support infrastructure that OpenAI and Anthropic have built over years. Performance on standardized benchmarks does not always translate to real-world reliability across diverse production use cases. The Hunter Alpha episode proved the model can compete on raw quality. Whether Xiaomi can compete on the surrounding infrastructure — support, reliability, uptime, and developer experience — remains to be seen.
