If you copy a Groq quickstart off a blog post, it almost certainly hands you this:
model="llama-3.3-70b-versatile"
That string doesn't work on a self-serve key anymore, and the failure is confusing because the model is still listed in the docs.
Open console.groq.com/docs/models and look at the columns, not the rows. Under Production Models, llama-3.3-70b-versatile and llama-3.1-8b-instant both carry an "Enterprise" badge, and the price and rate-limit cells read Contact Sales instead of a number. Every other production model has real figures there.
The rate limits page confirms it from the other side. The Free Plan table lists gpt-oss-120b, gpt-oss-20b, gpt-oss-safeguard-20b, qwen3.6-27b, qwen3.8-27b, compound, compound-mini, whisper, and the two orpheus voices. No Llama chat model appears on it. The only Meta entries left are llama-prompt-guard-2-22m and -86m — 22M and 86M parameter safety classifiers that score prompts and can't hold a conversation.
Those two classifiers are also, incidentally, where the famous "Groq gives you 14,400 requests/day" number lives. Chat models get 1,000 RPD.
The drop-in replacement is openai/gpt-oss-120b: same 131K context, and it's on the free tier.
I keep the current Groq free-tier numbers up to date here: Groq free API limits
Top comments (0)