Repos
llama-modchip claims DeepSeek V4 Flash, a 284B model, runs on a 64GB PC with no GPU. What the numbers say
A llama.cpp-based project says custom compression squeezes DeepSeek-V4-Flash into under 59 GiB of RAM at about 9–15 tokens a second on a Ryzen laptop chip. The figures are self-reported, the memory margin is razor-thin, and the files won't load in stock llama.cpp or Ollama.