Memory Workbench
Account for model, cache and runtime memory without making the unknown parts disappear.
Open toolModels / machines / measurements
Run it here. Find out what it can actually do.
Memory budgets, reproducible inference tests, and local vision experiments. Keep the model, the configuration, and the evidence in the same conversation.
Start with the first field notePut something on the bench
All toolsField notes
All field notesThe weights fit. The model doesn’t.
Estimate a bounded KV-cache allocation and keep it separate from weights, runtime buffers, and the machine you actually have.
Read the field noteTokens per second need footnotes
Compare inference runs without mixing prompt processing, generation, different models, and missing configuration.
Read the field noteA vision demo is not a measurement
Turn an impressive detection into a bounded test with known inputs, visible errors, and a reproducible environment.
Read the field noteThe confidence setting is not a measuring instrument.
Local AI becomes interesting when it does useful work under conditions you can inspect. Start with a bounded task, record the environment, and keep the failures in the results.
See where this connects to the other departments