Controls how many words each Chain of Draft step should target (2-20 words).
Chain of Draft (CoD) is configured to instruct the model to:
If enabled, adds a *brief* final self-check reflection (starting with "Reflection:") just before the CoD final answer. This is in addition to the three detailed reflections.
Controls randomness: lower values make responses more focused. (Fireworks default: 0.6)
Controls diversity via nucleus sampling. (Fireworks default: 1)
Restricts selection to the K most likely tokens. (Fireworks default: 40, 0 means disabled)
Reduces repetition of similar tokens.
Reduces repetition based on frequency.
Maximum length of the response. (Qwen3-235B context is larger, but completions are often capped)
Model Context Protocol (MCP) tools extend the model's capabilities by allowing it to request actions like:
Select which tools you want to make available to the model. The model will then decide if and when to use them by outputting a structured request.
Note: For this local test version, actual external API calls for tools (like to Tavily) are not performed by the browser. The UI will show if the model requests a tool.