Gemini 3 Flash

Gemini 3 Flash is the budget workhorse of the pack's text models: Google's small fast model, hosted by Runware and exposed as a node that takes messages and returns text. The whole point is that the LLM becomes a first-class citizen of your ComfyUI graph - prompt rewriting, captioning, JSON structuring, workflow decisions - without you standing up a separate LLM server or wiring a second API key into everything.

The real draw is the messages input, which comes from the Runware Messages builder: you build role/content pairs (system, user, assistant) and chain them, so this node does multi-turn conversations, not just one-shot completions. That turns it into a genuinely useful automation part: a system prompt that imposes a format, a user message that carries your input, and a text output that feeds a parser or a prompt processor downstream.

What you set

• messages (required) - the conversation, from the Messages builder. Each entry is role + content; chain them for multi-turn.
• images - optional IMAGE input, which makes this vision-capable: ask the model to describe or caption what's in a frame.
• settings.systemPrompt - the global instruction for behavior and output style.
• settings.maxTokens - up to 65536, default 4096.
• settings.temperature (0–2, default 1) and settings.topP (0.95) - the creativity dials.
• settings.thinkingLevel - minimal/low/medium/high, how much internal reasoning before answering.
• toolChoice - gated tool-calling: auto/any/tool/none, with a tool name when you force one.
• includeUsage - include token stats in the response.

Output is text, a STRING. advanced_json covers the gaps: extra audio/document/video inputs, stop sequences, tools.

How it works

Standard pack flow: taskType: textInference over REST through the Runware SDK, and the model's reply comes back as a string (joined across results). All the heavy lifting happens on Google's/Runware's side - no local model, no VRAM, cost on the title bar. Because it's a hosted call, each text out is a fresh inference; keep your context in the Messages chain, not in your head. For the price of a local setup you get a model that's been through a generation of improvements you'd never see by running the same checkpoint for months.

Installing

ComfyUI Manager → search Runware → install → restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/Runware/ComfyUI-Runware
pip install -r ComfyUI-Runware/requirements.txt

No model downloads; deps are runware-sdk, pillow, soundfile. API key via Settings → Runware API key, RUNWARE_API_KEY, or runware auth login.

Troubleshooting

Empty or truncated output usually means maxTokens is cutting the answer off - bump it. Model rambling off-format? Put the format contract in settings.systemPrompt, where it belongs, rather than begging in the user turn. If you need heavier reasoning for a hard task, the family scales up: Gemini 3.1 Flash Lite is the cheaper sibling, Gemini 3.1 Pro is the brainier one.