New AI Tech Makes Big Models Smaller Without Losing Smarts

New AI compression methods GSQ and RCO are releasing quantized versions of Qwen3.8-27B, promising higher-quality models at the same file size with both the search and quantizer learned. GSQ (Gumbel-Softmax Quantization) learns grid assignments and scales, closing most of the scalar-vs-vector gap at 2 to 3 bits, while staying deployable in GGUF. RCO (Riemannian Constrained Optimization) assigns a quantization type to every tensor under a strict size budget using gradient descent on the task loss, with no per-constraint tuning. The release includes three GGUFs (2.50 / 2.75 / 3.00 bpw, 8.4 to 10.1 GB) plus the vision projector, and runs unmodified in llama.cpp, Ollama, and LM Studio. According to the article, these are the best size-to-accuracy GGUFs available for Qwen3.8-27B at these file sizes, matching or beating the strongest published quantizations across every size evaluated. Specific results include a 3.00 bpw version matching the base model on AIME25 (100.00) and coming within about one point on GPQA-Diamond and LiveCodeBench, a 2.75 bpw version exceeding the BF16 zero-shot average (75.70 vs 74.34), and gains of +10.0 AIME25, +8.6 GPQA-Diamond, and +4.6 LiveCodeBench over UD-IQ2_S at a matched ~8.4 GB size. It's the first in a planned series of SOTA GGUFs from the lab.