On May 16, Xiaomi officially announced via its Weibo account that its large language model, MiLM, has completed its备案 (registration) process. This model will gradually be integrated into various Xiaomi products, including cars, smartphones, and smart home devices, and will be made available for more users to experience in the future.
The MiLM model first gained attention during last August's C-Eval and CMMLU large model evaluations, where it ranked 10th overall in C-Eval and secured the top position among models with similar parameters. According to its GitHub project page, MiLM-6B is a large-scale pre-trained language model developed by Xiaomi, boasting a parameter count of 6.4 billion.
C-Eval data reveals that MiLM-6B demonstrates high accuracy across 20 subjects in STEM (Science, Technology, Engineering, and Mathematics). This includes fields such as physics, chemistry, and biology. The model also performs well in humanities subjects, though it requires improvement in disciplines that involve "abstract thinking," such as law, mathematics, programming, and probability theory.
In assessments of social sciences, MiLM-6B achieved satisfactory accuracy in 8 out of 10 subjects, excluding "Education and Geography." Additionally, during evaluations in the humanities, the model showed commendable results in the "History and Law" domains.