Gemma 4 31B

ComfyUI isn't just for images anymore, and this node is a good demonstration of why. Gemma 4 31B is Google's open-weights reasoning model, and the Runware pack serves it as a text node: you build a conversation, it thinks, and out comes a string. Why would you want a 31B LLM inside a ComfyUI canvas? Because text nodes have become the plumbing of complex workflows - prompt rewriting, captioning, workflow logic, a helper agent that decides which branch to take. If you'd rather not run a 31B model locally (which on most home GPUs means slow or out-of-memory), this gives you the capability on demand, billed per request.

The thing to understand up front is that this is a cloud inference node, not a local loader. There's no GGUF to download, no KV-cache to manage. You hand the API a chat log and it returns text. The node's job is to turn that round-trip into something ComfyUI-friendly.

How it works

Text nodes in this pack work via a Runware Messages builder. The node's required messages socket takes a chain of message objects (role: user or assistant, plus content) - you stack them to build a conversation, then the node sends a textInference request and joins the results into a single text output. Note the advanced_json escape hatch on this node: it's merged raw into the request, and the tooltip names inputs.videos and settings.stopSequences as the things you can't reach any other way. So if you need a video fed to the model, or want to stop generation at a specific token sequence, that's where it goes.

The inputs that matter

• messages (required) - the conversation, via Runware Messages. settings.systemPrompt (a plain multiline string, separate from the builder) is where you set the model's overall behavior - style, role, constraints. For a prompt-rewriting workflow this is the most important widget on the node.
• settings.thinkingLevel - off or high (default high). This is a reasoning model; high gives it room to think before answering. Drop to off when you need fast, cheap, straight answers - for many pipeline tasks you don't want the reasoning overhead.
• images - an IMAGE input, so this is a vision-capable model too. Wire a generated image in and ask it to describe or critique it.
• settings.maxTokens - an off-by-default gate with a paired settings.maxTokens_value. Leave the gate off to use the model default; toggle it on if you need to cap output length.
• settings.temperature - likewise a gate (with temperature_value, 0–2). Off = model default, which is the sane choice for deterministic workflow logic. On = you're doing creative writing.
• settings.topK / settings.topP / settings.minP / the penalty sliders - sampler tuning for the brave. For pipeline use, leave them.
• numberResults (1–4) - multiple answers with different seeds. Handy when you want to pick the best of a few caption drafts.
• includeUsage - returns token-usage stats in the response if you want to keep an eye on spend.

The single output is text (a STRING). Wire it into any text-input socket in your graph - a prompt field, a filename builder, a switch.

Installing

ComfyUI Manager → search Runware → install → restart, or:

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

Then set your API key in ComfyUI Settings → Runware API key, via the RUNWARE_API_KEY env var, or runware auth login. Nothing to download - the model lives on Runware's side.

Where people trip up

The most common gotcha is expecting a ComfyUI-native image or hidden state out of a text node - it's a plain string, so if you're feeding it into an image pipeline you need an intermediate node that parses text. Second, remember the maxTokens gate is off by default: if the model's output is getting truncated for you and you need longer answers, flip the gate and raise the value, don't assume you have to live with it. And cost - every call is metered, and the title bar shows each run's price. A 31B reasoning model with thinkingLevel: high can chew through tokens; the includeUsage toggle is there so you can actually see it.