400 "Unable to access non-serverless model" (Together AI)
A request that worked yesterday now returns 400 Unable to access non-serverless model meta-llama/Llama-3.3-70B-Instruct-Turbo with a link to create a dedicated endpoint. Together removes models from its serverless pool on a schedule published at docs.together.ai/docs/deprecations, and since mid-2026 the list includes Llama 3.3 70B Turbo and DeepSeek V4 Flash 0731 (22 October), GPT-OSS 20B (15 September), DeepSeek V4 Pro and Kimi K2.7 Code (27 August) and Kimi K2.6 (19 August). The id keeps appearing in GET /v1/models with running: false, so a check on the model list does not catch it; the first sign is this error in production.
Last verified November 20, 2026 against Together deprecations, Together error codes · 5 min read
| Provider | Together AI |
| HTTP status | 400 Bad Request |
| Error type | invalid_request_error |
| Message | Unable to access non-serverless model <id>. Please visit https://api.together.ai/models/<id> to create and start a new dedicated endpoint for the model. |
| Variants | endpoint_not_configured (Endpoint has no deployments configured for routing) and endpoint_not_ready |
| When it happens | The model was removed from Together's serverless pool; it still appears in GET /v1/models with running: false |
| Can you wait it out? | No. Serverless removals are permanent; the model is only available on a paid dedicated endpoint |
- Not your fault, not temporary. The model left Together's serverless pool on a published date.
- Three ways out: a dedicated endpoint on Together (billed per GPU hour), a different serverless model on Together, or the same id on another host that still serves it.
- Check the list before you move: the deprecations page shows what else is scheduled.
What happened
Together runs two kinds of inference: a shared serverless pool billed per token, and dedicated endpoints billed per GPU hour. When demand for a serverless model drops, Together retires it from the pool and keeps it available only as a dedicated deployment. The removal history on the deprecations page runs to more than 150 ids since 2024, and the 2026 removals include some of the most referenced ids on GitHub.
| Together id | Removed from serverless |
|---|---|
| meta-llama/Llama-3.3-70B-Instruct-Turbo | 2026-10-22 |
| deepseek-ai/DeepSeek-V4-Flash-0731 | 2026-10-22 |
| openai/gpt-oss-20b | 2026-09-15 |
| google/gemma-4-31B-it | 2026-09-15 |
| deepseek-ai/DeepSeek-V4-Pro | 2026-08-27 |
| moonshotai/Kimi-K2.7-Code | 2026-08-27 |
| moonshotai/Kimi-K2.6 | 2026-08-19 |
| Qwen/Qwen3-235B-A22B-Instruct-2507-tput | 2026-07-10 |
| deepseek-ai/DeepSeek-R1, deepseek-ai/DeepSeek-V3.1 | 2026-05-14 |
| mistralai/Mixtral-8x7B-Instruct-v0.1 | 2026-04-16 |
| meta-llama/Meta-Llama-3.1-8B-Instruct-Turbo | 2026-03-06 |
| meta-llama/Llama-4-Scout-17B-16E-Instruct | 2026-02-06 |
The error text and the running: false flag are what you see; Together's page gives removal dates but no replacement for serverless retirements, so the choice of successor is yours.
Fix 1 — a dedicated endpoint on Together
The link in the error creates a dedicated deployment of the same model. It starts in minutes and is billed per GPU hour from the moment it runs, idle or not, so it suits steady heavy traffic and nothing else. Autoscaling to zero removes the idle cost at the price of cold starts.
Fix 2 — a serverless model Together still runs
Pick a current id from GET /v1/models (filter running: true) and change the model string. For Llama 3.3 70B there is no serverless Llama successor on Together at the time of writing; GPT-OSS 120B and the Qwen3.5 family are the usual picks. Expect different output, context limits and tool-calling behaviour, and re-test.
from together import Together
client = Together()
models = [m.id for m in client.models.list() if getattr(m, "running", True)]
print(models)Fix 3 — keep the id, change the base URL
The retired ids are open-weight models that other hosts still serve. Inference APIs serves Llama 3.3 70B Turbo, GPT-OSS 20B, Kimi K2.6, Kimi K2.7 Code and DeepSeek V4 Pro under exactly the Together ids, and accepts the other retired ids as aliases of the closest model it serves (DeepSeek-V4-Flash-0731 as DeepSeek V4 Flash, Mixtral and the Llama 3.1 Turbo ids as Llama 3.3 70B, QwQ and Qwen 2.5 as Qwen3.8 Flash). Billing is per token with no endpoint to rent.
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; $0.20 / $0.50 per 1M tokens
messages=[{"role": "user", "content": "Hello"}],
)The together SDK itself takes a base_url, so the same change works without switching libraries. The full alias table is on the Together alternative page.
Catching the next one before production does
Watch the deprecations page, not the model list. A retired serverless id stays listed with running: false, and a one-token test call per model per day is the only check that reproduces what users will hit. A router with a fallback (LiteLLM, or your own) turns the 400 into a retry on another host instead of an outage.
Frequently asked questions
Is the model gone from Together entirely?
No. It can be deployed as a dedicated endpoint, billed per GPU hour. Only the shared, per-token serverless option is gone.
Why does GET /v1/models still list it?
Together keeps retired ids in the list for dedicated-endpoint use; the running flag is false. Treat running: false as "will return 400 on serverless".
Does Together announce these in advance?
The deprecations page carries a short scheduled-removals table (the November 2026 rows were MiniMax M3, Inkling and Qwen3.8-2.4T), but the big 2026 removals appear in the history table with the removal date only.
Where Inference APIs fits
Fix 3 below keeps the model id and changes the base URL: Inference APIs serves Llama 3.3 70B Turbo, GPT-OSS 20B, Kimi K2.6, Kimi K2.7 Code and DeepSeek V4 Pro under the same ids, and accepts the other retired Together ids as aliases of the closest model it serves.
Something changed or wrong? Tell us and we will re-verify the entry.
