GPT-5.4 Mini

If you're wiring a language model into a workflow that runs often - rewriting every prompt before sampling, summarizing every caption, tagging every batch - you don't want the flagship model for every call. GPT-5.4 Mini is the same model family with the same control surface, tuned to be quicker and cheaper. It's the node you default to when the answer doesn't need to be brilliant, just right.

How it works

A textInference node hitting GPT-5.4 Mini through the Runware cloud. The required messages input takes a RUNWARE_MESSAGES socket from the Runware Messages builder (Runware/Params) - you compose the conversation there, this node just runs it.

The controls match the full 5.4 node, which is the point:

• settings.systemPrompt - the system instruction that shapes every response.
• settings.temperature (default 1, 0–2) and settings.topP (default 0.95) - the creativity dials are right here, so you can make output as deterministic or as loose as you like.
• settings.thinkingLevel - (default)/low/medium/high/xhigh. For a "just fix my prompt grammar" job, low is plenty and cheap.
• settings.maxTokens (default 4096, up to 128k) - response cap.
• images (IMAGE) - vision is available too: attach the image you're describing and the model will actually look at it.
• toolChoice / toolChoice.type / toolChoice.name with advanced_json for tool definitions, plus seed, numberResults, and includeUsage.

Output is a text STRING that feeds any text consumer in your graph.

Install

Part of ComfyUI-Runware:

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

Restart (or ComfyUI Manager → search "Runware"). Cloud model, no downloads. API key: Settings → Runware API key, RUNWARE_API_KEY, or runware auth login.

Gotchas

Per-token cost, same as any of these - the title bar shows it. The Mini is a genuinely good everyday model, but it's still smaller than 5.4; for long multi-step reasoning or anything that needs deep chains, keep the flagship or the Pro around and route the boring stuff here. And the statelessness rule applies to every text node in the pack: no built-in conversation memory, so multi-turn means wiring output back into the Messages builder.