a phone company beats a tech giant in ai with 12x less investment

meta has been researching ai since 2013.
xiaomi only started in 2016.

meta spent $57.4b on r&d in 2025.
xiaomi spent $4.8b – 12x less.

meta released its first llm in february 2023.
xiaomi released its first in april 2025 – two years later.

yet on 3 out of 4 independent ai benchmarks, xiaomi's mimo v2.5-pro just outscored meta's muse spark:

gdpval-aa: mimo 54% vs muse spark 46%
τ²-bench: mimo 94% vs muse spark 92%
long context reasoning: mimo 73% vs muse spark 70%

the only benchmark meta wins is terminal-bench hard – coding and agentic tasks – where muse spark scores 46% vs mimo's 43%.

more money. more experience. more employees. more years. and still, on most benchmarks, xiaomi's model is ahead. that's not luck. that's efficiency.