Local models · public library snapshot
Search the public Ollama library by capability, download size, context, and a VRAM budget. The Keep scrapes the library HTML itself. The counts on this page are recomputed in the browser and checked against evidence.json.
Method
The catalog is a static snapshot. scrape.py downloads ollama.com/library, then each family’s tags page, and writes data/models.json plus data/evidence.json. There is no Ollama API key and no account.
A tag is on the VRAM shortlist when it is not a cloud tag and its listed download size is at most 75% of the budget you pick. 16 GB therefore keeps tags at 12 GB or under. That leaves room for a KV cache. It is not a measurement of VRAM, and a model can still run out of memory at a long context.
Quant hints are read from the tag name. Most default tags, such as qwen3:14b, do not print a quant, so they stay unspecified. Pull counts are the rounded figures on the library cards.
For a worked hardware answer on a 16 GB RTX 5060 Ti, use the 5060 Ti Ollama lab. This catalog is the filter. That lab is the recommendation.