AtomicChat/DeepSeek-V4.1-Flash-NVFP4-nvidia
Text Generation • 763B • Updated • 365 • 5
DeepSeek V4.1 Flash
NVFP4 experts for Blackwell, calibrated by NVIDIA's recipe and measured against the original: checkpoint plus the metrics.
Note NVFP4 experts, calibrated (cnn_dailymail + Nemotron v2, 64x512). Equal to an uncalibrated cast within noise; W4A4 costs ~1.7 pt top-1, 0.8% ppl on text.
AtomicChat/DeepSeek-V4.1-Flash-NVFP4-metrics
Updated • 273 • 1Note KLD, top-1, perplexity on neutral/code/agentic for nvidia, flat and the reference against itself; coverage; amax dumps; every log.