SambaNova API Node

The short version

Most ComfyUI nodes spend your VRAM. This one spends an API key instead. The SambaNova API Node (SambaNovaNode) is a thin wrapper around SambaNova's hosted chat API: you type a prompt, it makes one HTTP request to a 70B+ model running in SambaNova's cloud, and hands the text response back to your graph as a string. No model files to download, no GPU headroom to find, no local LLM to babysit. It's the kind of node you reach for when you want an LLM in the workflow - writing or rewriting the prompt, generating a subject list, deciding something for you mid-graph - but you'd rather not run one locally.

That pattern used to be a browser tab. LLM-assisted prompting has since become a node, and hosted-API nodes like this are one flavor of it: cheap, fast, and completely free of your hardware.

How it works

The whole node is a single POST to https://api.sambanova.ai/v1/chat/completions - an OpenAI-compatible endpoint. The code builds a messages array from two inputs (a system_prompt and a prompt), sends it with your api_key as a Bearer token, then digs choices[0].message.content out of the JSON reply and returns it as a string. stream is hardcoded off, there's no chat history, no temperature, no token limit. One request in, one answer out, then it's done. Minimal is the point - this is a ~40-line node that exists to save you wiring raw HTTP yourself.

The inputs that matter

There are four, all plain strings, and honestly the only one that trips people up is the first:

• api_key - your SambaNova Cloud key (default is literally YOUR_SAMBANOVA_API_KEY, which is a 401 waiting to happen). Get one from SambaNova's site; the service runs on credits, not free forever. Where people get burned: the key is a plain workflow input, so it gets saved into your workflow JSON. Share a .json with this node in it and you've just handed someone a live billing account. Strip it before sharing, or keep such workflows private.
• prompt - the user message, multiline.
• system_prompt - the role. This is where you tell it "output comma-separated Danbooru tags" or "rewrite this as a structured prompt for a tag-based model" - the whole LLM-assisted prompting play.
• model - defaults to DeepSeek-V3-0324, one of the open-weight models SambaNova hosts. It must match their catalog exactly; their model list rotates as models ship and get retired.

The output

One output, response (STRING). Wire it into anything that takes a string - most usefully a CLIP Text Encode or the prompt box of a sampler, so the model's reply becomes the actual prompt for your image generation. It's text in, text out; there's no image output and no way to chain a conversation.

Install

Dead simple, and it's the nicest thing about this pack:

• ComfyUI Manager - search for comfyui-sambanova (published under the SambaNova category) and hit install.
• Or clone it manually:

  cd ComfyUI/custom_nodes
  git clone https://github.com/neggo/comfyui-sambanova

  Then restart ComfyUI.

The only dependency is requests - no torch, no transformers, no model download, nothing heavy. It installs in seconds on any machine, GPU or not.

Troubleshooting

• Your output starts with API Error: - that's the node's error-handling, not a crash. Every failure (bad key, wrong model name, network hiccup) is caught and returned as the response string instead of stopping the queue. That's convenient right up until a dead key silently feeds "API Error: 401 Unauthorized" into your image prompt and you get a garbage render with no error box. The full response is printed to the ComfyUI console, so check there when output looks wrong.
• Model not found - the hardcoded default name only works if SambaNova still serves that exact ID. Check their current model list and update the field.
• It's stateless and bare - no temperature, no history, no streaming. If you need actual chat or control over sampling, this node will feel thin fast; it's built for one-shot text-in-your-workflow, nothing more.

If you want a big hosted LLM riding along in your graph without touching your GPU, this is the one to grab - just don't share the workflow with the key in it.