deepseek-v4.1POST
DeepSeek V4.1 Flash 748B runs ~40 tok/s on stock llama.cpp
The author repacked a GSQ-RCO quant into upstream GGUF format and ran DeepSeek V4.1 Flash 748B on one RTX PRO 6000 plus RAM at about 40 tok/s, with no fork or third-party engine.
vveerrgg
@vveerrggArchived 14 hours agoOriginally posted 2026.10.28
vveerrgg @vveerrggOriginally posted 2026.10.28
@deepseek_ai V4.1 (748B) on stock llama.cpp. No Strata, no fork, no third-party engine. Repacked @ISTA_DASLab’s GSQ-RCO quant into upstream format: same weights, byte for byte. Runs on 1 RTX PRO 6000 + RAM: ~40 tok/s. 🤗 https://huggingface.co/HumanjavaEnterprises/DeepSeek-V4.1-Flash-1017-GSQ-RCO-upstream-GGUF 🛠️ https://github.com/HumanjavaEnterprises/deepseek-v41-upstream
x.com/vveerrgg/status/2115211511824298284Archived 2026.10.28 12:02