Groq Chat

Your GPU is busy diffusing, and the LLM that's supposed to improve your prompts is asking for another 20GB of VRAM. That's the classic ComfyUI squeeze: the graph is all about local inference, but a decent chat model is the one thing you can't fit. Groq Chat is the escape hatch. It's a node that sends your text to Groq's hosted models over the API and gets answers back at speeds that make a local Llama feel like it's running on a potato.

Groq is the inference provider known for absurd token throughput - its whole pitch is "extremely fast responses," and it's had a free API tier for years, which is why so many ComfyUI LLM packs have been built on it since 2024. The catch is that every call is a round trip to their servers, so you need an API key from console.groq.com and an internet connection. No key, no node.

What you'd actually use it for

The LLM-in-the-loop jobs that make ComfyUI workflows dramatically better: expanding a three-word idea into a real prompt, generating negative prompts, filling wildcards, rewriting captions before they hit the text encoder. Wire text into a Show Text node (or a text-loader node) and the output feeds straight into CLIPTextEncode. If you've ever seen a workflow that "runs everything through an LLM first," this is that node.

How it works

Under the hood it's one chat.completions.create call, but with a few ComfyUI-friendly touches. It builds the message array from three places - an optional conversation_history (a JSON array of prior messages), an optional system_prompt, and your current prompt - then sends it off with three retries and exponential backoff baked in. On success you get three outputs: text (the answer), usage_info (token counts, which matters if you're paying per token), and an updated conversation_history as JSON. Loop that last one back into the conversation_history input and you've got multi-turn chat without a single extra node. Neat trick.

The inputs that matter

You'll mostly touch these:

• prompt - your message. Multiline, so paste freely.
• model - defaults to llama-3.3-70b-versatile; the dropdown also has llama-3.1-8b-instant (the speed pick), the OpenAI gpt-oss-120b / gpt-oss-20b models, Groq's agentic groq/compound pair, and even meta-llama/llama-guard-4-12b. The model list is read from configs/models.json, so if Groq ships something new you can just add it there.
• temperature (0–2) and top_p (0–1) - the usual randomness knobs. Default temperature is 1; drop toward 0 when you want consistent structured output.
• max_tokens - cap on the response length, up to 32768.
• seed - set a non-negative value and Groq will try to make the output reproducible, if the model supports it.

Optional api_key and system_prompt round it out. Leave api_key empty and it falls back to the GROQ_API_KEY environment variable or a .env file.

Installing it

From ComfyUI Manager, search "Groq Nodes" and install, or do it by hand:

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

Then restart ComfyUI. The dependencies are light - groq, aiohttp, Pillow, numpy, python-dotenv - and there are no model files to download; everything runs on Groq's side. That's the whole point.

Common issues

The node returns error strings in text rather than failing loudly, so check the output: "Groq API key not found…" means you haven't set the key anywhere. The .env route works but only if load_dotenv() finds the file from wherever ComfyUI is launched, which can be surprising - the env var or pasting the key into the node is the more reliable setup. Expect occasional 429 rate-limit errors on the free tier, and remember this is a small, young pack: it runs arbitrary Python on your machine like any custom node (the ecosystem has a history there), so install from the real repo and keep it updated.