Loading the catalogue…
Loading the catalogue…
What a model can do at the API boundary — tool calling, vision input, extended thinking, context caching. Fifteen first-class capabilities across input, output, serving, and tools categories. Stav grades every model in the catalogue against this vocabulary.
What the model can ingest — text, images, audio, video, and structured data.
system_promptModel accepts a separate system message distinct from the user message.
vision_inputModel accepts image inputs alongside text.
audio_inputModel accepts audio input.
document_understandingModel accepts native document files (e.g. PDF) end-to-end.
What the model can return — streaming, structured output, tool calls, audio.
tool_callingModel emits structured function calls per OpenAI tool-use schema.
parallel_tool_callsModel can emit ≥2 tool calls in one response. Modifier on tool_calling.
structured_outputModel returns JSON conforming to a supplied JSON Schema (constrained decoding).
json_modeModel returns syntactically valid JSON when asked. Looser legacy capability.
Operational features — prompt caching, long context, batch inference.
Agentic features — function calling, code execution, retrieval, browsing.
web_search_nativeProvider offers a native web-search tool exposed through the same endpoint.
code_execution_nativeProvider offers a native code-execution tool.
file_search_nativeProvider offers a native file-search / RAG primitive.
Source · capability registry. Stav’s capability vocabulary is locked at v3.6 and shared across the public site, customer portal, and admin curation tooling. New capability slugs require a registry write and a curator-pipeline conformance pass before they land here.