Skip to content
Fantástico Mundo de Jon
RSS

I run AI models on my own hardware and publish the numbers.

Six machines, from 12 to 96 GB: four NVIDIA GPUs, Apple Silicon and unified memory. Lab

Writing

  1. Coloring Pages with Local AI: From Prompt to a File That Prints Well

    Generating the image is the easy part, and that's where most tutorials end. What determines whether the child can paint is what comes next: binarizing, vectorizing, and printing without jagged edges.

  2. Qwen3.6 27B: How Much VRAM It Really Needs

    A 27B model that uses half the KV cache of a Llama 3 8B. The hybrid architecture breaks the usual calculation — and decides, for less than 1 GB, which cards are left out.

  3. How much VRAM an LLM really uses

    The back-of-the-envelope math — parameters times bits — is off by several gigabytes, because it ignores the KV cache. Here is the full calculation, with the formula, the numbers and how to verify it on your own GPU.

  4. Measuring if a spec file makes the model perform better. It doesn't.

    In July, I argued that structured specs change the outcome. Measured in October: 360 generations, 3 local models, 3 versions of the same request. What matters is the written information, not the file format.

  5. The PRD comes first: the spec that prevents the agent from writing incorrect code

    The model doesn't make mistakes due to ignorance: it makes mistakes because it filled the gap the prompt left, and filled it in a plausible way. What changes in the code when the contract comes before the request.