Explorer
See where the memory goes.
These maps come from the published checkpoints' own metadata and tensor headers. Each cell is a weight matrix. Its colour shows the chosen format; its drawn size helps show the space it takes.
Choose a published model.
Violet marks more compressed weights, teal marks higher precision, and amber marks weights kept at their original precision. Hatched cells were fixed by the recipe rather than chosen by the allocator.
Published allocation
…
Download this map's data (JSON). Technical detail reveals format names, exact counts and byte accounting.
A map of choices, not a quality score.
Columns follow the model's layers. Rows group similar operations. An empty position means that operation is absent there. Switch to an even grid to give each decision the same space.
The footprint view fits row and column totals, with minimum heights for tiny rows. Individual cell areas are approximate. Hover or tap a cell for its measured bytes; the data link provides every assignment without using the canvas.
Explore a recorded budget sweep.
The budget slider replays recorded allocator solves on Qwen3-4B. Predicted loss is the cost model's estimate, not served quality. The activation lever is a separate illustration.
…
Illustration only. This control keeps weight bytes fixed and changes an activation-precision label in promotion order. It is not a measured quality or speed result, and is not the PrismaAQUA allocation method.
Look beneath the assignment.
The sensitivity view shows why some changes matter more than others. It is one input to a decision that also depends on size and the available formats.