NVIDIA ships NVFP4 DeepSeek-V4.1-Flash, 81GB smaller with near-identical quality
AccBalanced · x · 2026-10-31
NVIDIA released an official NVFP4 quantization of DeepSeek-V4.1-Flash-1017, shrinking the weights by about 81GB (475→400 GiB).
Benchmarks show near-identical quality: Terminal-Bench 2.1 drops only from 81.81 to 81.11 (-0.70 pts) and SciCode from 57.58 to 57.32 (-0.26 pts). Weights are on Hugging Face (nvidia/DeepSeek-V4.1-Flash-1017-NVFP4), tested with vLLM at max reasoning effort; the original checkpoint uses MXFP4 experts.
Related event: NVIDIA Releases NVFP4 Quantized DeepSeek-V4.1-Flash(3 posts)→
More from Infra
- CXMT validates G6 DRAM engineering samples in Q3, mass production expected within a year — zephyr_z9 · 2026-10-31
- AI scarcity moved to atoms: memory, GPU and wafer margins hit 58%-80% — sujingshen · 2026-10-31
- Anthropic's S-1 risk factors: compute tops the list at 13.4%, edging out regulation — ayushtweetshere · 2026-10-31
- TU Eindhoven shows glass interposer with BGA below and MPO optical port on top — jwt0625 · 2026-10-31
- AI Boom Triggers Massive PCB Shortage With Orders Booked Until 2030 — chemist_slime · 2026-10-31
- Beff Jezos doubles down on Cerebras: GPU backends feel 'stupid' for frontier-scale fast inference — beffjezos · 2026-10-31