qwen3-235b-topiary-stream-kit

qwen3-235b-topiary-stream-kit — build kit + an honest negative

This is NOT a servable model on 24 GB Apple Silicon, and this card says so with measurements. It is the build kit for the day you (or we) have a 48–64 GB machine: skeleton checkpoint (4.5 GB, expert-free), universal 2D salience floor (11.8 GB, k=256), routed-salience orders, and the plane manifest. The 122 GB of expert plane memmaps are NOT uploaded — rebuild them in ~1 h from the community checkpoint:

python src/split.py --src mlx-community/Qwen3-235B-A22B-Instruct-2507-4bit-DWQ \
    --out artifacts/qwen235-stream --layout full-memmap --consume
# then drop this kit's files (floor, orders, README) into the artifact dir

The measured verdict at 24 GB (four serving modes, six designs)

Three converging walls, all measured: pool coverage (11–16% of experts ≪ the ~30–60-experts/layer working set), floor budget below the width cliff, and flat per-expert salience (a 17% prefix captures only 53.5% of energy — vs ~94% on concentrated models). No pool size fits: C=20 already peaks at 22 GB. Token-level retry-on-miss does not converge (L·k·P(miss) ≫ 1 against load-balanced expert tails; falsified live at 99% retries).

Falsifiable prediction: at 48–64 GB (pool coverage ≈50% + this kit's floor as the safety tier) the same runtime serves this model with nothing new built. If you own such a machine, this kit is a one-hour experiment.

Provenance

Skeleton and floor built with split.py / floor.py; salience orders profiled through the exact-mode pager (the source checkpoint had been consumed — the memmaps are the model; 3,072 calibration tokens). Full story: https://github.com/jesusluque/topiary-stream