OrcaSAQ-2 27B is a sensitivity-aware mixed-precision quantized checkpoint derived from Qwen3.8-27B, released under Apache-2.0 on Hugging Face. It reduces the original ~54 GB BF16 weights to 12.3 GB while retaining 93.2% top-1 token agreement, near-identical perplexity, 262K context, tool calling, and thinking mode. Designed for coding, terminal, browser, and multi-tool agent workflows on modest hardware such as 16 GB GPUs.
