orcarouter/OrcaSAQ-2-27B — new model trending #30 on Hugging Face
OrcaRouter released OrcaSAQ2 27B, a 12.3 GB 3-bit quant of Qwen3.8-27B with near-BF16 perplexity.
OrcaRouter released OrcaSAQ2 27B, a sensitivity-aware mixed-precision quantization of Qwen3.8-27B that shrinks a 54 GB BF16 checkpoint to 12.3 GB at about 3.21 bits per weight. On WikiText-2 it reports perplexity 5.6482 versus 5.6468 for BF16, a 0.02% increase, plus 93.2% top-1 agreement and 0.031 mean KLD. The Apache-2.0 checkpoint keeps a 262K context window, thinking mode, and tool calling, and is served with vLLM. The card lists 70.0 on SWE-bench Verified and 58.4 on Terminal-Bench 2.1, noting those figures are not strict model-only comparisons.
46