Leak: DeepSeek V4.1 Flash Is Actually a 718B Sparse Model

Leaks claim DeepSeek V4.1 Flash is a 718B sparse model (500GB on disk), with an unverified report of a 196B-parameter lookup table dubbed "Engram" replacing computation with table lookups.

2026-09-10 ~ 2026-09-11 · 2 related posts