Bad Theory Labs Releases BTL-3: A 27B Agentic Model Compressed to 8.39GB
Bad Theory Labs has released BTL-3, a 27B parameter open-weight model optimized for agentic coding and tool use. Its "Compact" edition is compressed to just 8.39GB (under 2.5 bits per parameter) using a custom quantization stack, allowing it to run entirely locally. This release demonstrates that large-scale models can be heavily compressed to a size smaller than an 8B model in fp16 while retaining over 92% of their original performance. It enables developers to run highly capable, private, and local coding agents on consumer-grade hardware. The model features a 262K context window and achieves a 95.12% pass@1 on HumanEval and 88.5% on the BFCL v4 AST benchmark. The custom quantization stack uses techniques like packed AVQ2 decoder tensors, affine INT4, and rank-32 output correction to minimize degradation.
## BACKGROUND
Quantization is a technique used to reduce the memory footprint of LLMs by representing weights with fewer bits, though standard methods often degrade performance at very low bitrates. GGUF is a popular binary file format optimized for fast loading and running models locally, while benchmarks like the Berkeley Function Calling Leaderboard (BFCL) evaluate a model's ability to call external tools and APIs.