Young Tries AI

Everything on DeepSeek-V4-Flash

1 mention, newest first.

  1. llama-modchip repo lists a 284B DeepSeek model running on 64GB RAM, no GPU

    The GitHub repo llama-modchip, shared on Hacker News, lists DeepSeek-V4-Flash 284B checkpoints of 56.93 GiB and 56.40 GiB. On a 64 GB desktop with no GPU, the README lists 8.75–9.55 tok/s, and 11.2–25.3 tok/s with speculative decoding. A 30B Nemotron VQ3 GGUF is listed for 16–32 GB laptops at 22.4 tok/s.

    Partly usable