A "phone" company is now competing with Anthropic on AI benchmarks. Xiaomi's MiMo-V2-Pro ranks #3 globally on agent tasks.
Phone maker Xiaomi secretly released a top-tier AI agent that developers mistook for DeepSeek V4.
Xiaomi, the Chinese electronics giant known for smartphones, has made a stunning entry into the frontier AI race with its MiMo-V2 model family. The flagship MiMo-V2-Pro, a 1-trillion parameter model with 42B active parameters, secretly spent a week on OpenRouter under the codename 'Hunter Alpha.' The developer community, impressed by its performance, widely assumed it was the unreleased DeepSeek V4. Xiaomi's subsequent reveal that it was their own model sent shockwaves through the AI community, demonstrating that the company, led by a researcher from DeepSeek, has built world-class AI capabilities virtually overnight.
The technical specs are formidable. MiMo-V2-Pro achieves agent benchmark performance comparable to Anthropic's Claude Opus 4.6 but at a fraction of the cost ($1/$3 per million tokens vs. $5/$25). More impressively, the fully open-source MiMo-V2-Flash (309B params) scores 73.4% on the SWE-Bench coding benchmark, edging out Claude Sonnet 3.5 (72.8%) while costing just $0.10 per million tokens—30 times cheaper. Xiaomi also released the multimodal MiMo-V2-Omni for processing text, image, video, and audio, and MiMo-V2-TTS for expressive speech, creating an integrated agent stack where Pro thinks, Omni perceives, and TTS speaks.
This launch fundamentally reshapes the competitive landscape. A year ago, Xiaomi was synonymous with affordable hardware. Today, it fields a four-model AI family that directly challenges established frontier labs like Anthropic on both performance and price. The MiMo-V2 release signals the intense velocity and surprising origins of innovation within the Chinese AI ecosystem, proving that deep AI research is no longer confined to dedicated AI startups or Western tech giants.
- MiMo-V2-Pro (1T params) ranks #3 on agent benchmarks, matching Claude Opus 4.6's performance at 1/8th the API cost.
- MiMo-V2-Flash is fully open-source, scores 73.4% on SWE-Bench, and outperforms Claude Sonnet while being 30x cheaper per token.
- The models were stealth-tested as 'Hunter Alpha,' fooling the community into thinking it was DeepSeek V4 before Xiaomi's surprise reveal.
Why It Matters
A major hardware company now competes with top AI labs on performance and cost, dramatically increasing accessible, high-quality AI.