A compact model Yantra-1B has been released for invoking tools

Developers presented Yantra-1B — a specialized language model with 1 billion parameters, optimized for tasks of invoking external functions and tools, the solution is focused on working in conditions of limited computational resources. The model demonstrates a PAS 0.6775 score on the ToolACE-300 benchmark. According to the test results, the performance of the new model exceeds the base model's performance by 3.52 times. High accuracy is achieved due to the combination of a lightweight router and fine-tuning of the MiniCPM5-1B architecture using the QLoRA method. The technical implementation of the model is focused on generating arguments for API calls. The use of QLoRA allows for efficient adaptation of the model's weights to specific tool invocation tasks while maintaining a minimal amount of occupied memory, which is critical for deployment on peripheral devices. The release of Yantra-1B expands the possibilities for creating autonomous agents that can interact with external software without the need to use heavy LLMs. The model is available for integration into projects that require high query processing speed with a limited token budget.