Inference APIs
Reference/Errors/Together

400 "Unable to access non-serverless model meta-llama/Llama-3.3-70B-Instruct-Turbo" (Together AI)

If your app pinned meta-llama/Llama-3.3-70B-Instruct-Turbo on Together, the most referenced Together id on GitHub (about 18,000 code files mention it with Together), it started returning 400 Unable to access non-serverless model on 2026-10-22. Together removed the model from its shared serverless pool; the id still appears in GET /v1/models with running: false, and the only way to run it on Together now is a dedicated endpoint billed per GPU hour. The request, the key and the SDK are all fine.

Last verified November 20, 2026 against Together deprecations, Together error codes · 4 min read

ProviderTogether AI
HTTP status400 Bad Request
Error typeinvalid_request_error
MessageUnable to access non-serverless model meta-llama/Llama-3.3-70B-Instruct-Turbo. Please visit https://api.together.ai/models/meta-llama/Llama-3.3-70B-Instruct-Turbo to create and start a new dedicated endpoint for the model.
Removed from serverless2026-10-22
Together's listed replacementnone for serverless
Can you wait it out?No. The removal is permanent; the model is only available as a paid dedicated endpoint
Short answer
  • Permanent: meta-llama/Llama-3.3-70B-Instruct-Turbo left Together's serverless pool on 2026-10-22.
  • On Together: a dedicated endpoint (per GPU hour) or a different serverless model.
  • Elsewhere: the same id is served as the same model on Inference APIs; change the base URL only.

What happened

Together's deprecations page lists meta-llama/Llama-3.3-70B-Instruct-Turbo in its serverless removal history with the date 2026-10-22 and no replacement. The model remains deployable as a dedicated endpoint, which is what the error's link offers. Retired ids stay in the model list, so a health check that only lists models keeps passing while every real request fails.

First, find where the id is pinned

The id is often in more places than the one that errored: a default in a settings file, an environment variable on the server, a LiteLLM or LangChain config, a notebook. Grep the repository and the deployment environment before changing anything, or the next deploy brings the error back.

Find every pin
grep -rn "meta-llama/Llama-3.3-70B-Instruct-Turbo" . --include="*.py" --include="*.ts" --include="*.js" --include="*.yaml" --include="*.yml" --include="*.toml" --include="*.env*" --include="*.json"

Option 1 — a dedicated endpoint on Together

Follow the link in the error. The endpoint starts in minutes and is billed per GPU hour while it runs, with or without traffic, so it pays off only for steady heavy load. Autoscaling to zero removes idle cost at the price of cold starts of a minute or more.

Option 2 — a serverless model Together still runs

There is no serverless Llama successor on Together; GPT-OSS 120B is the usual pick, with different output and tool-calling behaviour. Changing the model changes the output, the context limit and sometimes tool-calling behaviour, so re-run your evals.

Option 3 — keep the id, change the base URL

Inference APIs serves meta-llama/Llama-3.3-70B-Instruct-Turbo under the same id, as the same open-weight model, at $0.20 / $0.50 per 1M tokens. Billing is per token from a prepaid balance; there is no endpoint to rent and no daily cap.

Python (OpenAI SDK)
from openai import OpenAI
import os

client = OpenAI(
    base_url="https://api.inferenceapis.com/v1",   # was https://api.together.xyz/v1
    api_key=os.environ["INFERENCEAPIS_API_KEY"],
)
resp = client.chat.completions.create(
    model="meta-llama/Llama-3.3-70B-Instruct-Turbo",   # unchanged
    messages=[{"role": "user", "content": "Hello"}],
)
Python (together SDK)
from together import Together

client = Together(base_url="https://api.inferenceapis.com/v1", api_key=os.environ["INFERENCEAPIS_API_KEY"])
resp = client.chat.completions.create(model="meta-llama/Llama-3.3-70B-Instruct-Turbo", messages=[{"role": "user", "content": "Hello"}])

Streaming, tool calling and JSON mode behave as on Together. The Together alternative page lists every retired id and what answers it here.

Frequently asked questions

Will Together bring the serverless model back?

Its deprecations page has no record of a serverless model returning. Plan as if it will not.

Does the Together SDK work against another base URL?

Yes. Both the Python and Node together packages accept base_url; the OpenAI SDK works too.

How do I avoid the next one?

Watch docs.together.ai/docs/deprecations rather than the model list, run a one-token test per model daily, and put a fallback host behind a router so a 400 becomes a retry rather than an outage.

Where Inference APIs fits

Option 3 above keeps the id and changes the base URL: Inference APIs serves meta-llama/Llama-3.3-70B-Instruct-Turbo as the same model at $0.20 / $0.50 per 1M tokens.

Something changed or wrong? Tell us and we will re-verify the entry.