Groq LLM

Groq's whole identity is speed - they run open-weight LLMs on custom hardware so fast that responses feel instant. This node puts that inside ComfyUI as a text-generation node: give it a prompt, get a string back, in a fraction of the time an equivalent local model would take. If your workflow needs a fast language-model step - rewriting a prompt on every render, generating variations for a batch - Groq is the low-latency pick of this pack's two LLM nodes.

The other reason to use it: Groq's free tier is genuinely generous, so this is the cheapest "brain in the workflow" option you'll find. It's the prompt-rewriting workhorse, not the deep-thinking one.

How it works

The node calls Groq's OpenAI-compatible endpoint (api.groq.com/openai/v1) using the OpenAI Python client pointed at Groq's base URL, and returns the completion text as a STRING. Optional image input goes along for vision-capable models. reasoning_effort is passed through for the models that support thinking (like the gpt-oss line), and disabled disables it entirely.

The inputs that matter

• prompt - the instruction. For prompt-rewriting, this is where you'd put "rewrite this for FLUX, keep the subject, add cinematic lighting."
• model - eight choices: llama-3.3-70b-versatile (the default, a strong all-rounder), llama-3.1-8b-instant (small/fast), openai/gpt-oss-120b and openai/gpt-oss-20b (OpenAI's open-weights reasoning models, hosted by Groq - the 120b is the one to pick for serious rewriting), meta-llama/llama-4-scout-17b-16e-instruct, qwen/qwen3-32b, and the compound models groq/compound / groq/compound-mini. The default is fine to start; gpt-oss-120b is the quality jump.
• max_completion_tokens - default 8192, up to 131072. Enough headroom for long rewrites.
• temperature (default 1.0) - the pack's LLM nodes run hot by default; for stable prompt-rewriting, drop it to 0.5 or lower.
• reasoning_effort - disabled (default), default, low, medium, high. Flip this up for harder reasoning tasks, not for every call - it costs latency.
• system_instruction (optional) - role-setting, e.g. "you are an expert SDXL prompt writer."
• image (optional) - attach a reference image for vision-capable models, turning this into an image-caption or image-query node.
• seed / api_key - key in the field or .env (GROQ_API_KEY).

The output

response - a STRING, wired into any text input downstream.

Installing it

One of ~20 nodes in the Aryan185/ComfyUI-ExternalAPI-Helpers pack:

cd ComfyUI/custom_nodes
git clone https://github.com/Aryan185/ComfyUI-ExternalAPI-Helpers.git
cd ComfyUI-ExternalAPI-Helpers
pip install -r requirements.txt

Restart ComfyUI; find "Groq LLM" under text/generation. Needs the openai package. Get a free Groq key at console.groq.com, paste it in or set GROQ_API_KEY in .env.

Where people get burned

Not every model in the list supports every feature. Set reasoning_effort to high on a model that doesn't do reasoning and you can get API errors or ignored settings - check Groq's docs for which models take it. Also, some of these (especially the compound and gpt-oss models) are served under preview/experimental programs and can be rate-limited or swapped out; if the model enum points at something no longer live, you'll get a 404-ish error that looks like a node bug. And remember temperature: at the default 1.0, re-runs of the same prompt drift. That's fine for creative variation, annoying for captioning.

The honest comparison: use this for speed and free-tier - prompt rewriting, batch variation, quick summarization. If you need the smartest possible answer regardless of cost, the pack's OpenAI LLM node is the one.