MiniMax M2.7

Easily the most surprising node in the Runware pack, because it isn't image or video at all: it's a hosted LLM. MiniMax M2.7 is a large language model, exposed as a chat node that takes a RUNWARE_MESSAGES conversation in and returns generated text. If you've been treating ComfyUI as purely a diffusion playground, this is the pack reminding you that a generation graph can host a text model just as easily as a diffusion model - and that the same cloud account and API key power all of it.

The practical pitch: you can now do prompt engineering inside the graph. Generate a prompt with an LLM node and feed the string into a positivePrompt socket. Write a negative prompt programmatically. Summarize a caption, batch-rewrite a dataset's descriptions, or drive a captioning loop - all without leaving ComfyUI and without installing a local model. For that job, a hosted M2.7 is wildly more convenient than running a 70B-class model on your own machine.

Inputs that matter

The required messages input takes a RUNWARE_MESSAGES value from the Runware Messages builder (also under Runware/Params), which has just two fields: role (system/user/assistant) and content (the text). Build your conversation there and wire it in.

• settings.systemPrompt - system-level instruction that steers behavior for the whole generation. This is where you say "you are a prompt engineer who writes photorealistic Stable Diffusion prompts."
• settings.temperature - randomness, 0–1, default 1. Lower for deterministic rewrites, higher for creative variety.
• settings.topP - nucleus sampling, default 0.95. Leave it alone unless you know the difference.
• settings.maxTokens - response ceiling, default 32768, up to 196608. The tooltip's range tells you M2.7 is a serious context-length model.
• toolChoice / toolChoice.type - tool-calling strategy (auto, any, tool, none), with toolChoice.name when you pin one tool. The advanced_json field exists for the tools array itself - so yes, you can build a tool-using agent in a ComfyUI graph.
• includeUsage - include token usage stats in the response. Turn it on if you're watching your budget, which you should be.

seed, numberResults (up to 4), and outputFormat (a one-choice TEXT dropdown) are the rest.

How it works

The node sends minimax:m2.7@0 as a textInference task through the Runware SDK. The text output is a plain STRING - which is the whole trick: a string wires anywhere. Straight into another node's prompt field, into a text consumer, or into the pack's own image nodes. Cost prints on the title bar.

Install and gotchas

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

Restart; API key via Settings → "Runware API key", RUNWARE_API_KEY, or runware auth login. Paid per token-generation run.

The trap is forgetting the messages builder exists and expecting a text box on the node itself. There isn't one - you must wire a Runware Messages node in, role and content per message. Second trap: a hosted LLM can burn through a top-up fast if you leave maxTokens at 32768 and forget includeUsage; for prompt-rewriting work, set a low max and watch the cost line.