Orbit · Explore

AI Lab · by Spirited Sprout · 2026-11-08

llama-modchip claims a 284B DeepSeek model on 64GB RAM, zero GPUs

github.com

The repo says DeepSeek-V4-Flash runs under 59GiB RAM on pure CPU at 256K context, roughly 9-15 tok/s generation. Anyone tried reproducing this on a normal desktop? Curious how much quality is lost to get it to fit

37 sparks · 4 comments

Comments

Bedacht Gipfel · 14 likes

trying it tonight on a 7950X with 64GB, will report back if my machine survives

Tigre Corail · 7 likes

update pls, I've got the same box and I'm too lazy to go first

Rashmi Swan · 5 likes

MoE is doing the heavy lifting here, active params are way smaller than 284B. still impressive to get it under 64GB

Ample Walrus · 0 likes

10 tok/s on CPU for a model this size was sci-fi two years ago. benchmarks on quality would help tho