DEV Community

AI Tech Connect
AI Tech Connect

Posted on Originally published at aitechconnect.in

NVFP4 Lands: 119B Open Weights Now Fit in 66 GiB

Originally published on AI Tech Connect.

What actually changed This is not a model release. Mistral Small 4 shipped on 16 March 2026. The weights are six months old. Only the NVFP4 checkpoint activity is recent. The official NVFP4 checkpoint exists. mistralai/Mistral-Small-4-119B-2603-NVFP4 is published by the Mistral organisation itself, built with llm-compressor in collaboration with the vLLM and Red Hat teams, and lands at roughly 66 GiB. Most Qwen NVFP4 checkpoints are not from Qwen. The Qwen3.8-27B-NVFP4 repositories verified on the Hub around 19 September 2026 come largely from third parties — RadixArk, QUASAR-QAT, Ostfralla. That is a supply-chain fact, not a footnote. Recipes differ, so results differ. Two uploaders can both say "NVFP4" and ship materially different artefacts. The licence does not move. Quantisation does…


Read the full article on AI Tech Connect →

Top comments (0)