Gemini 3.1 Flash-Lite

If the Gemini family in this pack is a lineup, Flash-Lite is the workhorse: cheap, fast, good enough for the jobs that don't need the flagship. As an LLM node inside ComfyUI, it's the one you wire up when you want a text-generation step that won't dent your credit balance - prompt expansion, tag translation, batch rewriting, generating variations for a loop. It's the hosted Google model behind a BizyAir call, so there's zero local footprint: no model file, no VRAM, just an API key and a network connection.

Inputs and outputs

• system_prompt - required. The framing and rules.
• user_prompt - required. The actual request.
• temperature - 0 to 2, default 1.
• max_tokens - default 32768, up to 65536. Generous for long rewrites.

Optional: enable_thinking (default false on this tier - it's a lite model; thinking is available but off by default, and leaving it off is exactly why Flash-Lite is cheap), enable_search (default true), plus the pack-wide api_config override and skip_error.

Outputs: result (STRING) and request_id. The result needs a text preview node to see it - a bare graph won't display the string for you.

How it works

The usual BizyAir machinery: serialize system/user prompt plus sampling settings, submit to the gemini-3.1-flash-lite-official/large-language-models endpoint, poll until the completion lands, return the text. Your BizyAir key from settings or BIZYAIR_API_KEY pays for it; api_config overrides base URL/key for that one request.

The honest take

Flash-Lite is the node you reach for when you don't need a showcase. Prompt expansion doesn't need the flagship; neither does translating a prompt from English to Japanese, or churning out ten variations for a batch loop. The trap is forgetting the defaults: enable_search being on means every call can hit the live web, which is slower and costlier than a plain completion - flip it off for pure language work. And if you ever find yourself asking "why is this returning mush," the answer is usually that you left thinking off and search on for a job that wanted the opposite.

On failure with skip_error on, you get the pack's placeholder output (for text nodes, the error string) instead of a crash - check the console for the real message.

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, and wire result into your prompt pipeline. For the price, it's the node you'll leave in your graph and forget about - which is exactly what a workhorse should be.