Quantized models published only if they clear a stated accuracy bar vs bf16. Measured deltas per model card. NVFP4 A16: Ada and newer, no Blackwell.
UIST Labs
company
Verified
AI & ML interests
LLM quantization & model compression · NVFP4 / weight-only W4A16 · accuracy-gated certification (publish only on pass vs bf16) · efficient inference beyond Blackwell (Ada /Ampere / Hopper, FP4 Marlin) · reproducible, DOI-cited releases · reasoning & instruction-tuned models