The model occupies about 18 GB of video memory and can be run on a single server accelerator.
The MWS AI team has presented the multimodal model Cotype Light 3, designed for integration into AI agents. The model is capable of simultaneously processing text and visual data, such as contracts, drawings, and images. This allows agents to solve multi-stage tasks without switching between systems.
Cotype Light 3 contains 9 billion parameters, which ensures performance with low resource requirements. As the developers promise, it works on standard corporate equipment and quickly learns specific tasks, reducing infrastructure costs and speeding up implementation.
Denis Filippov, CEO of MWS AI, emphasized that a compact specialized model is cheaper to operate and works more accurately in a specific domain than a universal system with a redundant number of parameters.
The model occupies about 18 GB of video memory and can be run on a single server accelerator. It is compatible with domestic hardware and software systems, including the PAK Skala^r Machine AI. MWS AI trains Cotype Light 3 and other models on the cloud capacities of MWS Cloud.
