Bad Theory Labs posts BTL-3, a 27B agentic coding LoRA on Qwen

A small independent lab called Bad Theory Labs has quietly posted a model card on Hugging Face for BTL-3, a 27 billion parameter agentic coding model that is technically a LoRA adapter sitting on top of Qwen/Qwen3.6-27B, and the numbers on that card are the kind of thing worth pausing on. According to the page, BTL-3 hits 95.12% on HumanEval pass@1 in thinking mode, 88.1% on LiveCodeBench v6, and 88.5% on the full BFCL v4 AST tool-calling set, including a 91.2% irrelevance score for correctly declining to call tools when it shouldn't.

The reason a release like this is interesting is the shape of it. This is not a from-scratch pretrain. It is a PEFT LoRA with rank 32 and alpha 64, roughly 934 MB on disk, released under Apache-2.0, with a Compact edition the card claims runs under 2.5 bits per parameter at 8.39 GB while retaining 92.2% of tool-use behaviors. If those figures reproduce, the practical implication is that a team with a single well-specced GPU could self-host a coding agent that on paper sits in the neighborhood of hosted assistants for the specific slice of work these benchmarks measure.

For teams that don't train models in-house, agentic coding has mostly been a paid-API story, and the interesting cost line for engineering leaders is what happens when a serviceable open substitute lands. An Apache-2.0 license plus a LoRA form factor means anyone can pull it, fine-tune it further on internal repos, and run it behind a firewall without touching the base weights.

The model card is currently the only source for any of this. Bad Theory Labs is not a well-known name, the page shows only 77 downloads in the last month, and the reporting doesn't give you a SWE-bench score, an independent replication, or any real disclosure of the training data or RL setup beyond a note that the maximum RL sequence length was 65,536 tokens. Treat the numbers as claims from the lab, not verified results.

Still, the release fits a broader pattern. Small labs shipping fine-tuned adapters that credibly claim frontier-adjacent tool-use scores on commodity open base models is the shape open source coding infrastructure has been converging on, and BTL-3 is a data point worth revisiting once someone outside the lab runs the evals.