According to a report by MarkTechPost, a new open-weight model named Underdog Saluki 27B has been released under the Apache 2.0 license. The model is a 2-bit GGUF quantization of Qwen3.8-27B, coming in at a compact file size of 7.89 gigabytes. In benchmark evaluations, this heavily compressed version outperforms the original 54-gigabyte base model specifically in tool-calling tasks, though it sacrifices performance in competition-level mathematics and general reasoning.
Understanding Extreme Quantization
Model quantization is a technique used to reduce the memory footprint and computational requirements of large language models by storing weights with fewer bits of precision. While standard deployments often use 4-bit or 8-bit quantization, pushing a model down to 2 bits like Saluki 27B drastically shrinks storage needs—allowing much larger models to run on consumer-grade hardware—but typically introduces degradation in complex reasoning and logic benchmarks. The fact that tool-calling accuracy has improved despite this extreme compression is a notable technical outcome.
Unanswered Questions and What to Watch
Many details regarding the development of Saluki 27B remain unknown. The announcement did not specify the exact calibration datasets used during the quantization process, nor did it detail the specific hardware setups utilized for testing. Furthermore, it is unclear how the model performs on broader conversational benchmarks outside of tool calling and math. Observers will need to watch for community evaluations and technical write-ups from developers testing the 7.89 GB file in real-world deployment scenarios.
As compressed open-source models continue to evolve, experiments like Saluki 27B demonstrate that trade-offs in capability are not always linear across all use cases.
Lead image: Nicolas Foster / Pexels
