DeepSeek V4.1 Flash (IQ2_XXS)

antirez's 2-bit GGUF of DeepSeek V4.1 Flash 1017, the October 2026 update to V4.1 Flash with large gains in reasoning, coding and agentic work. It is a 552B-parameter Mixture-of-Experts (8B active per token while reading a prompt, 16B while writing) with a 1M-token context and 196B parameters of Engram memory looked up from disk. The file is 341 GiB: 152 GiB of model weights plus 189 GiB of Engram tables that are never loaded into memory, so it needs a fast local SSD with room to spare. On a 128 GB Mac or DGX Spark the routed experts stream from the SSD, which makes it slower than models that fit in memory. Text only in Gezel for now. Produced for antirez's ds4 inference engine.

Gezel downloads this model the first time a gezel needs it, checks the file against its published hash, and applies tuned settings for it.

The model’s own page · Every model · Model scorecard