Developer hails DeepSeek efficiency: KV Cache shrinks 437x in three years

Developer gnukeith argues DeepSeek is severely underrated in engineering: V4.1 Flash runs 1M-token context in under 0.9GB of memory versus 320GB for Llama 3.3 70B. He notes KV Cache size has shrunk 437x in three years and that DeepSeek's quantization fits a 763B model into 305GB nearly losslessly.

2026-10-29 ~ 2026-10-29 · 3 related posts

• DeepSeek V4.1 Flash runs 1M context in under 0.9GB RAM vs 320GB for Llama 70B — gnukeith · 2026-10-29
• KV cache per token shrank 437x in 3 years — why this dev thinks AI gets cheaper — gnukeith · 2026-10-29
• Running a 763B-parameter model in ~305GB RAM near losslessly, thanks to DeepSeek's paper — gnukeith · 2026-10-29