AR6

Finds / notebook

Local AI workshop

What runs well on an Apple Silicon Mac, and what to test next.

7 finds, all kept, none read yet m · reasons proposed by a model on 13 Oct, waiting to be confirmed i

Question: What can local models do on a personal machine?

  1. post @akshay_pachaar on X. An inference engine four times faster than llama.cpp — @akshay_pachaar on X

    To test: an inference engine said to be four times faster than llama.cpp.i

    1. det. 13 Oct · interpretedReason proposed by a model from the content and Julien’s profile; not confirmed.

    re-check by 13 Nov · now interpreted

    kept, not read

  2. repo TensorFold Studio for Apple Silicon

    To test: text to image to video on MLX, in one command.i

    1. det. 13 Oct · interpretedReason proposed by a model from the content and Julien’s profile; not confirmed.

    re-check by 13 Nov · now interpreted

    kept, not read

  3. post Ivan Fioravanti on X. Qwen3.8-Flash-Next tool-calling showdown — Ivan Fioravanti on X

    To follow: tool calling compared across 9 setups and 5 inference engines.i

    1. det. 13 Oct · interpretedReason proposed by a model from the content and Julien’s profile; not confirmed.

    re-check by 13 Nov · now interpreted

    kept, not read

  4. article JetBrains on Hugging Face. Qwen3.8 and 3.6 27B blend — JetBrains on Hugging Face

    To follow: a blend of two Qwen models, with MLX weights.i

    1. det. 13 Oct · interpretedReason proposed by a model from the content and Julien’s profile; not confirmed.

    re-check by 13 Nov · now interpreted

    kept, not read

  5. repo TensorFold

    To test: fast, exact decoding on MLX behind an OpenAI-compatible endpoint.i

    1. det. 13 Oct · interpretedReason proposed by a model from the content and Julien’s profile; not confirmed.

    re-check by 13 Nov · now interpreted

    kept, not read

  6. post David Hendrickson on X. GGUF files usable directly on Apple Silicon — David Hendrickson on X

    To follow: loading GGUF directly on Apple Silicon.i

    1. det. 13 Oct · interpretedReason proposed by a model from the content and Julien’s profile; not confirmed.

    re-check by 13 Nov · now interpreted

    kept, not read

  7. article Glance Speedlab: a local VLM latency study

    To follow: PyTorch on MPS against MLX 8-bit, measured on a Mac.i

    1. det. 13 Oct · interpretedReason proposed by a model from the content and Julien’s profile; not confirmed.

    re-check by 13 Nov · now interpreted

    kept, not read