Groq retired llama-3.3-70b-versatile. Here is what broke, and three ways to fix it

On 16 August 2026 Groq shut down llama-3.3-70b-versatile and llama-3.1-8b-instant for free and developer tiers. A month later, hundreds of thousands of code files still pin the id. What the error looks like, who it hit, and the fix ladder.

Taylor Hawkes · September 16, 2026

If your app started answering every request with 404 The model `llama-3.3-70b-versatile` does not exist or you do not have access to it in the second half of August, this is why: Groq announced on 17 June and retired the model on 16 August for everyone except enterprise contracts. Its recommended replacement is openai/gpt-oss-120b (or qwen/qwen3.6-27b).

How widespread it is

We looked at GitHub in mid-September. Code search finds roughly 287,000 files that pin llama-3.3-70b-versatile against a Groq base URL, and about 2,200 issues and pull requests mentioning the id since the start of August — the migration scramble. Reading a sample of the affected repositories, the pattern is consistent: hobby and small production apps (chat assistants, hiring tools, e-commerce helpers, daily-brief cron jobs, Home Assistant voice pipelines) that picked Groq for the free tier and hard-coded the model.

Three fixes, in order of effort

Change the model string on Groq. openai/gpt-oss-120b keeps you on the free tier. It is a different model: better at reasoning and tools, and it thinks before answering, so raise max_tokens. Prompts tuned for Llama may need adjustment. If you also hit Groq's daily caps, note that the Developer-tier upgrade has been paused; see the entry on that.
Keep the exact model on another provider. Llama 3.3 70B is open-weight and still served elsewhere. On Inference APIs the Groq id is accepted as an alias, so the change is the base URL and key:

Python · same model id, different base URL

from openai import OpenAI
import os

client = OpenAI(
    base_url="https://api.inferenceapis.com/v1",   # was https://api.groq.com/openai/v1
    api_key=os.environ["INFERENCE_API_KEY"],
)
resp = client.chat.completions.create(model="llama-3.3-70b-versatile", messages=[{"role": "user", "content": "hello"}])

Stop pinning a single provider. The retirement was the third in a year for that family of ids on Groq (llama3-70b-8192 → llama-3.3-70b-versatile → gone). Put the model id in configuration, add a fallback provider for 404 and 429, and watch the deprecations tracker.

A note on honesty

We serve Llama 3.3 70B as the FP8 "Turbo" variant, at $1.35 per million tokens — more than GPT-OSS 120B costs here ($0.20 / $0.80). If you are not tied to Llama's exact behaviour, GPT-OSS 120B is the better deal on both providers. The reason to keep Llama is prompts, tests and outputs you have already tuned; that is a real reason, and it is the case we built the alias for.