27B (dense)
262,144 tokens
Apache 2.0 model artifacts; MIT native runtime
Open-weight agentic coding and structured tool-use family derived from the Qwen3.6-27B hybrid-attention causal language model. The maximum-quality BTL-3 release is the frozen RL-0013 checkpoint distributed as a 933,974,032-byte rank-32, alpha-64 PEFT LoRA adapter for a pinned Qwen3.6-27B base revision, with a 65,536-token maximum reinforcement-learning sequence length and a 32,768-token launch benchmark context. It is trained for repository work, coding, debugging, structured single and parallel function calls, multi-turn execution, recovery after failed actions or verifier feedback, and abstaining when no tool call is appropriate. Bad Theory Labs reports 95.12% HumanEval pass@1, 88.1% on its completed 193-case LiveCodeBench v6 run, 88.5% BFCL v4 AST across 1,240 cases, and 26.35% on BigCodeBench-Hard Instruct; its BFCL results include 95.5% for multiple calls, 87.0% for parallel calls, and 91.2% irrelevance. The coding benchmark results used thinking mode, but the current supported release default is non-thinking because the RL-0013 reasoning policy can become repetitive or fail to terminate. The grouped release also includes BTL-3 Compact, a complete 64-layer standalone text model packaged in a single 8.39 GB GGUF at under 2.5 effective artifact bits per parameter. Compact combines packed AVQ2 and UniSVQ decoder tensors, affine INT4 tensors, selected higher-precision behavior-sensitive islands, packed vocabulary matrices, a rank-32 output correction, and a compact behavior adapter without loading or reconstructing the BF16 base. Its July 24 behavior refresh reports 98.4% conditional retention on an internal 63-case teacher-correct tool-contract gate. The included native runtime directly executes the packed representation on CUDA and Metal and provides an OpenAI-compatible server plus custom LM Studio and Ollama CLI integrations; stock GGUF engines do not directly decode its AVQ2 representation.