2 pointsby bhamm-lab7 hours ago1 comment
  • bhamm-lab7 hours ago
    Wrote up some notes from upgrading my local llama.cpp setup on AMD Strix Halo hardware and testing a batch of newer open models.

    Main findings:

    - Kimi Linear 48B works well as a generalist on my hardware (fast, consistent).

    - Qwen3 Coder Next is my new default for coding tasks.

    - Aggressive quants (Q2_K_XL) on massive models (200B+) can still be useful for long-running/background tasks, even if they aren't interactive.

    Happy to answer questions about the dual AI Max+ 395 setup or how I run models in kubernetes.