BTL-3
A 27B coding and tool-use language model family for agentic repository work, released both as a full-precision PEFT adapter on a Qwen3.6 base and as a standalone, packed-weight GGUF edition for native local inference.
BTL-3 is a 27-billion-parameter model trained for agentic coding and structured tool use: reasoning, acting, inspecting tool results, recovering from failures, and stopping when no action is needed, including explicit abstention when a tool call is unnecessary and multi-turn recovery after failed actions. It reports 88.5 percent on the BFCL v4 AST tool-calling benchmark and 95.1 percent on HumanEval.
The release has two editions: BTL-3, a rank-32 PEFT adapter that requires the Qwen3.6-27B base model and runs through Transformers or vLLM for maximum quality on servers; and BTL-3 Compact, a standalone 8.39 GB GGUF file using mixed low-bit quantization that runs through a packed llama.cpp runtime with no base model required, aimed at Macs, workstations and private, offline agents.
This repository provides the native runtime, integrations and reproducible packaging for both editions; the model weights themselves are hosted on Hugging Face.