AnalysisAI ModelsOctober 27, 2026

ISTA-DASLab's 2.25bpw DeepSeek-V4.1-Flash quant tested against official API

Read original source →reddit.com

A LocalLLaMA user ran ISTA-DASLab's 2.25bpw GSQ-RCO GGUF quant of DeepSeek-V4.1-Flash-1017 for several days on Project-maya v1.1.27 and reports it holding up well. The low average bit-width prompted the comparison against the full model via the official API.

1 source

More stories today

Open the live feed