Open-source models have been launched 7 and 32 billionth

Today we have released two large language models in Russian: T-Pro. 32 Billion parameters and updated T-Lite 7 billion parameters. They are based on Qwen models. 2.5 And they were trained in Russian.

T-Pro took second place on the benchmark MERA among all models, including proprietary ones, and the first place among open ones. T-Lite became the best Russian-language open-source model in the class 10 "Billions of parameters" on a number of industrial benchmarks.

The models can be downloaded from huggingface.

More details - very soon ;)