AI BriefingBTL-3 27B Agentic Model Released
Key point
Bad Theory Labs has released BTL-3, a 27B open-weight agentic coding model that runs from a single 8.39GB file, under the Apache-2.0 license.
Details
Bad Theory Labs has unveiled BTL-3, a 27B open-weight model specialized for agentic coding and structured tool use.
BTL-3 Compact compresses the entire model into a single 8.39GB GGUF file using less than 2.5-bit parameter compression. It is smaller than an fp16 8B model, while reportedly retaining 92.2% of the performance of the original 27B model.
Key Benchmarks (BTL-3)
- HumanEval: 95.12% pass@1
- BFCL v4 AST (all 1,240 cases): 88.5%
- Multi-tool calling: 95.5%
- Tool-call abstention: 91.2%
- Context length: 262K
For compression, the company applied its own in-house stack, including packed AVQ2 decoder tensors, affine INT4, precision islands, rank-32 output correction, and behavioral repair. It recorded a generation speed of 43 tok/s on an RTX PRO 6000 and supports fully local execution.
The model is licensed under Apache-2.0 and the runtime under MIT, and both are available on Hugging Face and GitHub.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.