The Models page

Four tabs, one table. frontend/src/features/recipes/recipes-content/catalog-table-shell.tsx holds the primitives — frame, header cell, group row, data row, numeric cell, hover action — and Recommended, Search Hugging Face, Your servers, and Downloads all draw on them. A padding or alignment change belongs in the shell, so all four tabs move together; a change made in one tab is a bug in the making.

Two rules in that page are easy to break by accident, so they are written down.

Weights may take at most 70% of the pool

FIT_BUDGET_RATIO in model-fit.ts. The rest of the machine's aggregate memory pays for the KV cache, activations, CUDA graphs, and the runtime itself, so a checkpoint that exactly equals the pool cannot actually be served. Every fit verdict in the app measures against poolGb * FIT_BUDGET_RATIO, and the curated catalog's sizes are chosen against it.

A model that does not fit is dimmed, never reddened. The eye should land on what this machine can run, not on what is broken.

Sizes come from the repository when the catalog records one, and fall back to total_params_b × bytes-per-param otherwise. The per-format constants live in BYTES_PER_PARAM; q4 is 0.6, not 0.5, because Q4_K_M keeps more bits in the attention tensors and lands near 4.8 bits/weight in practice.

The newer catalog wins

frontend/src/lib/api/studio.ts compares the controller's /studio/model-index against the copy bundled into the app and serves whichever carries the later updated date. The app and the controller deploy separately, so without this a freshly updated app shows stale picks until someone restarts the controller — and restarting the controller kills the running model. Ties go to the controller, since that is where an operator's curated override lives.

Intelligence scores are the Artificial Analysis Intelligence Index, attributed under the table through intelligence_source. A model AA has not benchmarked carries null, which renders as "not rated" — never as zero, which would sort it against models that genuinely scored badly.