pollinations

The columns

Each name links to the call this endpoint's verdict was read from.

Tools

We sent a tools array and the model came back with a well-formed tool_calls entry an agent could dispatch. A 200 with no tool call in it is not tool support, and is recorded as a no.

Schema

The strict structured-output mode: response_format: {"type": "json_schema", …}. The reply parsed and matched the schema we sent.

JSON

The weaker mode: response_format: {"type": "json_object"}. Valid JSON, with no shape enforced. The two come apart in both directions — OVH's Mistral-7B-Instruct-v0.3 does the strict form and emits invalid JSON for the weak one; llm7's meta-Llama-3.1-8B-Instruct-Turbo does json_object while json_schema answers 405.

Vision

A real image went in inline and a correct answer about it came back. We ask about something only a model that looked could know.

Draws

On an endpoint's own page: bytes came back with an image content type, and they really are an image. Not a 200 taken on trust.

The cell states

yes

A real call proved it, at the time the endpoint page prints beside that call. The dot's tooltip says the same thing where there is a pointer to hover with.

no

A clean negative: the parameter was refused, or the answer did not conform.

no, claimed

The negative contradicts the provider's own published claim — their listing advertises the feature and our call could not get it out of the model. This is the most interesting cell on the site and the reason the shelf exists.

never probed

No verdict. Either nobody has asked yet, or the check ran and got no usable answer back — Pollinations sits here because its anonymous pool answered 402 to every feature probe. It is a question, never a no.

The same columns, in the API

curl 'https://stillworks.supercapybara.com/api/llm/up?feature=tools,json_schema'

The names are the API's spelling of the columns: tools, json_schema, json_object, vision, image_out. Several are and, never or — a row comes back only if a real call proved every feature named. Models we never probed for one of them are left out rather than optimistically included, so a filtered shelf is shorter than a hopeful one.