BTL-3: A 27B open-weight agent model for agentic coding and structural tool use (huggingface.co)

BTL-3, a newly released 27 billion parameter open-weight model, specializes in agentic coding, structured tool use, and executing complex multi-turn tasks. It boasts impressive performance metrics, achieving a 95.1% pass rate on the HumanEval benchmark and demonstrating strong structured tool use capabilities across various scenarios. The model's architecture allows for a substantial context window of 262,144 tokens, facilitating in-depth reasoning and tool interaction during coding tasks. Notably, BTL-3 Compact provides a streamlined version of the model, packing full functionality into an 8.39 GB file—smaller than many existing models—while retaining a high behavior retention rate of 92.2%. This release is significant for the AI/ML community as it sets a new standard for operational efficiency and performance in coding agents—enabling applications such as debugging, structured function calling, and test-driven development. With open weights under the Apache-2.0 license and compatibility with various frameworks, BTL-3 can be easily deployed in both private and public environments. Its versatility allows for improved coding efficiency and reliability, making it a valuable resource for developers and researchers focused on intelligent tool use and automation in software development.