AI Radar
Support
LiveUpdated 2026-09-23 22:05 UTC

Vision tower on DeepSeek-V4-Flash-Vision-Exp is 0.41B params. The whole model is 304.65B.

Vision tower on DeepSeek-V4-Flash-Vision-Exp is 0.41B params. The whole model is 304.65B. 0.135 percent. The eyes are…

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.

Vision tower on DeepSeek-V4-Flash-Vision-Exp is 0.41B params. The whole model is 304.65B. 0.135 percent. The eyes are under a GB even at bf16. 97.3 percent of the checkpoint is routed experts, shipped as native fp4. Dense parts fp8. So a 304B model lands at 167.8GB on disk, 4.41 bits per weight before anyone re-quants it. That is the local wall. 167.8GB does not fit a 128GB Strix Halo UMA, and that is before a single token of KV cache. Unsloth has a GGUF out, have not run it yet, so no tok/s claim from me. Config math only. Config also lists one KV head, alternating compress ratios of 4 a

Posted by 코지베어 🐻 CozyBear (1.2k followers) 1 h ago · 0 likes · 5 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. 45.3k posts from 5k X accounts over the last 14 days, 1.9k tools, 19 markets. Collected every 5 minutes, fully re-ranked every hour — last update 2026-09-23 22:05 UTC. Full method.