Gemini 3.5 Flash

Look at the input list on this node and you'll notice something the other Gemini LLM nodes don't give you: top_p, top_k, seed, and stop_sequences. The 3.5 Flash LLM node is the one with actual sampling knobs, which makes it the text node you reach for when you want deterministic, controllable completions - not just "send prompt, get text." If you're building a prompt pipeline that has to reproduce its output, or a loop where generation must stop at a marker, this is the one.

It's the hosted Gemini 3.5 Flash, billed through your BizyAir key like the rest of the family. Flash-class speed and price, with a schema that exposes the levers power users actually pull.

Inputs and outputs

• prompt - required. The task.
• system_instruction - optional. The framing and rules. "You are a prompt engineer" goes here.
• stop_sequences - optional. Strings that halt generation when the model emits them. Great for forcing JSON-ish or list-structured output to end cleanly.
• seed - optional, default 0. Here's the reproducibility win: same seed, same prompt, same settings, same completion. The other nodes in this family can't promise that.
• temperature - 0 to 2, default 1.
• top_p - 0 to 1, default 0 (0 meaning "let the model decide," not "no sampling" - treat it as the nucleus knob you can leave alone).
• top_k - default 0, same convention.
• max_output_tokens - default 32768, up to 65536.

Outputs: result (STRING) and request_id.

How it works

The standard BizyAir API flow: serialize prompt, system instruction, and all the sampling settings, submit to the gemini-3.5-flash-official/large-language-models endpoint, poll, return the text. api_config overrides base URL and key per request; skip_error swaps a crash for a placeholder on failure.

The honest take

Where this node fits in the lineup: Flash-Lite is the cheap workhorse, Pro is the flagship, and 3.5 Flash is the controllable one. The seed and stop_sequences inputs are the whole story - they make text generation composable in a way the one-shot nodes aren't. If you've ever rerun a workflow and gotten a different prompt back for the same input, you know why reproducibility matters here. Use the seed for deterministic batches; use stop_sequences when the output has to terminate at a structural boundary so your downstream parsing never chokes.

One trap: top_p and top_k default to 0 with the "0 = unset" convention. Don't read that as "no randomness" - if you want tighter sampling, set them explicitly rather than assuming 0 means the model is being conservative.

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, confirm BizyAirPlus is ON and keyed, and set your seed. The first time a workflow reruns and produces the identical prompt, you'll get why this node exists.