Benchmarks & research
Piccolo is AgentScope's family of open 4B, 9B and 27B decision models that weigh candidate options in context and return probabilities, aimed at reviewing an agent's tool calls, choosing what context to keep and filtering data.
Only you can see your notes.
Screenshot unavailable. Open the experiment ↗
AgentScope reports 92.21% on JevBench Public for Piccolo-27B, plus 84.41% on HarnessJevBench Hard and 82.60% on DataJevBench, two benchmarks it built itself. Inside QwenPaw, a local Piccolo can approve harmless tool calls, flag destructive ones and help compact long conversations. The local runtime serves one model at a time, and an ordinary chat endpoint will not do: the decision path needs Piccolo's own /v1/systemone endpoint.
You can have a developer use Piccolo to check an AI assistant's planned computer actions before they run. In AgentScope's QwenPaw demo, Piccolo approved two harmless file reads and flagged an attempt to overwrite a file as destructive.
Your developer runs Piccolo on your own machine, for example through QwenPaw's local runtime, which serves one model at a time. An ordinary chat service cannot ask it these questions. Piccolo can still misjudge an action, so keep a person in charge of approving risky ones.