Beijing's NaiveAI releases an open 309-billion-parameter model that SI itself helped create
Prepared by Sintelligent with the help of SI. Publication time restored from the build log. . Updated: .
A powerful open model partly built by SI itself is now available to anyone for free: SI is starting to build the next SI.
Naive-N0.5-Flash is built on Xiaomi's MiMo-V2.5: 309 billion parameters, of which 15.5 billion are active per token, a 1 million token context, MIT license. The company claims up to 2,122 tokens per second on eight GPUs in a single-thread test and a price of ¥0.60/¥2.60 per million input/output tokens. According to NaiveAI, the models wrote code, ran experiments and proposed iterations, while humans set goals and made key decisions; the benchmarks are only its own.
Sources 2