Bad Theory Labs Open-Sources BTL-3, a 27B Agent Model for Agentic Coding
Key Takeaways
- ▸Achieves state-of-the-art performance on coding benchmarks: 95.1% HumanEval, 88.5% BFCL v4 AST, 88.1% LiveCodeBench
- ▸Compact 8.39 GB edition enables local inference on consumer hardware while retaining 92.2% of full model's tool-behavior capabilities
- ▸Specialized for structured tool use with 91.2% correctness rate on tool relevance detection; supports single, multiple, and parallel tool calls
Summary
Bad Theory Labs has open-sourced BTL-3, a 27B parameter agent model specifically designed for agentic coding, repository work, and complex multi-turn tool use. Built on Qwen3.6-27B and fine-tuned with reinforcement learning, BTL-3 demonstrates strong performance across multiple coding benchmarks: 95.1% on HumanEval, 88.5% on BFCL v4 AST, and 88.1% on LiveCodeBench (193-case run).
The model is released in two editions—the full version and BTL-3 Compact, an 8.39 GB package optimized for local deployment that maintains 92.2% of the full model's tool-behavior performance. Both are available as open-weight models under the Apache-2.0 license on HuggingFace, deployable via Transformers or vLLM with support for structured function calling and parallel tool workflows.
BTL-3 is engineered for structured tool use, long-context reasoning (262,144-token context window), failure recovery, and multi-turn execution. The model excels at deciding when tools are necessary (91.2% BFCL irrelevance rate) and handling complex workflows. Use cases span coding agents, repository inspection and repair, test-driven debugging, terminal-based task execution, and private or self-hosted agent deployments.
- Open-source Apache-2.0 release with reproducible frozen checkpoints and full thinking-mode coding support via Qwen3.6 architecture
- Designed for long multi-turn agent workflows with built-in failure recovery and 262,144-token context window
