Google Gemini 3.6 Flash (Official Edition)

This is the newest text node in the pack, and it reads like a checklist of things power users actually complained the earlier Gemini nodes lacked. One-million-token input window. response_mime_type and response_schema for structured JSON output. max_output_tokens you can push to 65,536. Hosted Gemini 3.6 Flash, billed through your BizyAir key, and it's the one text node here that can genuinely act as a data-processing step, not just a prompt expander.

The million-token window is the headline. That's the difference between "summarize a paragraph" and "here's a whole project's worth of notes, structure it." Most ComfyUI users won't need a million tokens - but the headroom changes what you'd even attempt.

Inputs and outputs

• prompt - required. Up to 1,048,576 input tokens.
• system_instruction - optional. The framing.
• stop_sequences - optional, up to 5 strings that halt generation.
• response_mime_type - combo, default text/plain. Set it to application/json and the model will emit JSON.
• response_schema - optional. A JSON schema for structured output. Here's the key interaction: the tooltip says it plainly - the schema is only valid when response_mime_type is application/json. Set the mime type first or the schema does nothing.
• seed - default 0 (-1 = random), for reproducible completions.
• max_output_tokens - default 8192, up to 65,536.

Outputs: result (STRING) and request_id.

How it works

Standard BizyAir flow to the gemini-3.6-flash-official/large-language-models endpoint: serialize prompt, instruction, stop sequences, and the response format, poll, return the text. api_config overrides base URL and key per request; skip_error returns the error text as the result instead of crashing.

The honest take

The structured-output story is the real reason to pick this over the 3.5 node. If you're feeding the result into downstream parsing - say, a workflow that turns a model's answer into workflow parameters - then text/plain answers are a constant parsing headache, and JSON-with-a-schema makes it deterministic. The combo of response_mime_type plus response_schema is exactly the pair you need, and the tooltip's warning is the gotcha everyone hits: change the mime type to application/json or the schema is ignored.

One more thing worth saying: default max_output_tokens here is 8192, not 32768 like the older nodes. If you're asking for long structured output and it truncates, that's the knob - raise it before you blame the model.

Install

With the pack:

cd /path/to/ComfyUI/custom_nodes
git clone https://github.com/siliconflow/BizyAirPlus.git
cd BizyAirPlus
python -m pip install -r requirements.txt

Restart ComfyUI, make sure BizyAirPlus is ON and keyed. Then feed it a project's worth of text and ask for JSON - it's the node that finally treats ComfyUI's text step like part of a real pipeline.