Guanaco 33B

Guanaco 33B is the 33B member of the Guanaco family released with the QLoRA paper in May 2023. Tim Dettmers and collaborators showed that a 33B model could be instruction-tuned in 4-bit quantization on a single GPU in about 24 hours, with the OASST1-tuned Guanaco variant reaching near ChatGPT-level performance on the Vicuna benchmark.

The adapter weights are released under the Apache 2.0 license, but using them requires the underlying LLaMA base weights and therefore compliance with the LLaMA license. QLoRA's memory-efficient fine-tuning approach went on to power a huge share of community fine-tuning work.