AI Radar
Support
LiveUpdated 2026-09-19 19:27 UTC

Opened the shard headers on both DeepSeek V4.1 Flash repos today. Native 510.3 GB, the…

Opened the shard headers on both DeepSeek V4.1 Flash repos today. Native 510.3 GB, the official NVFP4 build 527.3 GB.…

This is a AI post classified by Jev as Open & local models (a model release), kept by the AI Radar because it carries real work, not commentary.

Opened the shard headers on both DeepSeek V4.1 Flash repos today. Native 510.3 GB, the official NVFP4 build 527.3 GB. The quant is 17 GB heavier than the file it was made from. Same tensor names, same shapes, same packing. layers.0.ffn.experts.0.w1.weight is 2304 by 2560 bytes in both, two FP4 values per byte. The scales are the whole difference. DeepSeek stores one 8-bit exponent scale per 32 weights, the NVIDIA build stores one per 16. So 4.25 bits per weight becomes 4.5. 384 experts across 40 layers is 543.58B weights, and that extra quarter bit alone costs 16.99 GB. The measured gap is 1

Posted by 코지베어 🐻 CozyBear (1.2k followers) 1 days ago · 0 likes · 13 views · view the original post on X. Kept by the AI Radar as Open & local models.

More AI work like this

Every post is read and classified by Jev (TypeSafe): what it is, which market it belongs to, and whether the link is a real tool. 27.3k posts from 4.8k X accounts over the last 14 days, 1.2k tools, 19 markets. Collected every 5 minutes, fully re-ranked every hour — last update 2026-09-19 19:27 UTC. Full method.