Research

Technical notes from YogiQ AI Labs.

Short technical write-ups of work we did and measured. Numbers come with their settings and error bars, comparisons are re-run on one machine, and what did not work is reported next to what did.

  1. Technical note

    6 November 2026

    Fitting GLM-5.3-Flash on one DGX Spark: measured bit placement

    How YogiQ quantized Z.ai's GLM-5.3-Flash to a 108.3 GiB GGUF that fits one DGX Spark, by measuring block-by-block sensitivity and giving 3-bit expert weights to the two most sensitive blocks.

    Read the note