deepseek-v4.1POST

V4.1 Flash IQ2XS on a 7900 XTX: 10 tps decode, 800 tps prefill

Running DeepSeek V4.1 Flash 1017 at GSQ-RCO IQ2XS quantization on a 7900 XTX with 60GB DDR5 yields 10 tps decode and 800 tps prefill, with 200k context under an FP4 KV cache.

dextune

@dextuneArchived 2 days agoOriginally posted 2026.10.26

dextune @dextuneOriginally posted 2026.10.26

Open original on X

DeepSeek V4.1 Flash 1017 GSQ-RCO IQ2XS w/ 7900 XTX 7800X3D CPU 60GB DDR5 RAM Decode 10 tps Prefill 800 tps Context is 200k with an FP4 KV cache. - If only it had just a bit more VRAM or RAM. https://x.com/Chromadera/status/2114691152842498483

This post quotes another post: x.com/chromadera/status/2114691152842498483

x.com/dextune/status/2114701121528979954Archived 2026.10.27 12:10