Inference APIs

404 model_not_found: qwen/qwen3-32b (Groq)

If requests for Qwen 3 32B on Groq started failing with 404 The model `qwen/qwen3-32b` does not exist or you do not have access to it, the model was retired, not your key. Groq shut it down for free and developer-tier accounts on 17 July 2026, a month after the notice. It was the only Qwen model on Groq's production list at the time, and it had itself been the replacement for qwen-qwq-32b and mistral-saba-24b a year earlier.

Last verified November 18, 2026 against Groq deprecations, Groq model list · 5 min read

ProviderGroq
HTTP status404 Not Found
Error codemodel_not_found
MessageThe model `qwen/qwen3-32b` does not exist or you do not have access to it.
Announced / shut down17 June 2026 / 17 July 2026
When it happensSince 2026-07-17, on any free or developer-tier request that names this model
Can you wait it out?No — the model was retired; the request fails until you change the model or the provider
Short answer
  • Fastest: change the model string to openai/gpt-oss-120b on Groq and raise max_tokens, because it reasons before answering. Different model, so re-check your prompts.
  • Least code change: keep qwen/qwen3-32b and change only the base URL; Inference APIs accepts the id and answers with Qwen3.8 Flash.
  • So it does not happen again: read the model id from configuration and add a fallback for 404 and 429.

What happened

Groq's deprecations page lists the notice: announced 17 June 2026, shut down 17 July 2026, free and developer tiers only. The recommended replacement is openai/gpt-oss-120b; the notice also pointed at meta-llama/llama-4-scout-17b-16e-instruct, which Groq retired on the same date. Groq's model list has since carried qwen/qwen3.6-27b and then qwen/qwen3.8-27b, so a current Qwen id exists there, but it is a different model and not the one your prompts were tuned on.

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 second call for summarisation or titles, a framework default.

Find every copy of the id
grep -rn -E "qwen/qwen3-32b" . --include='*.py' --include='*.ts' --include='*.js' --include='*.json' --include='*.yaml' --include='*.yml' --include='*.toml' --include='*.env*'
# and the places grep cannot see: environment variables on the server, a hosted secrets store, the provider dashboard's default model

Option 1 — switch to Groq's recommended replacement

One string changes and you stay on Groq's free tier. What is different about openai/gpt-oss-120b:

  • Both models reason before answering, but Qwen 3 let you switch thinking off with enable_thinking: false; GPT-OSS cannot be switched off, only turned down with reasoning_effort: "low". Code that relied on a no-thinking fast path needs a larger max_tokens.
  • Output style differs. Qwen's multilingual and structured-output habits are not the same as GPT-OSS's; re-run your test cases, especially non-English prompts.
  • Tool calling and JSON output work on both.
One string changes
resp = client.chat.completions.create(
    model="openai/gpt-oss-120b",   # was "qwen/qwen3-32b"
    max_tokens=800,                 # reasoning counts against it
    messages=[{"role": "user", "content": "Hello"}],
)

Option 2 — keep the old id, change the base URL

Qwen 3 32B is an open-weight model but Groq's id was Groq's own. Inference APIs accepts qwen/qwen3-32b (and qwen3-32b, qwen-qwq-32b, qwen/qwen3.6-27b) as aliases and serves them with Qwen3.8 Flash, the current generation of the same family: a 1,000,000-token window, thinking on by default with the same off switch (reasoning_effort: "none"), tool calling and JSON output, at $0.15 / $0.50 per million tokens with cached input at $0.02.

Base URL and key change, model string stays
from openai import OpenAI

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="qwen/qwen3-32b",   # the retired Groq id, accepted as an alias
    messages=[{"role": "user", "content": "Hello"}],
)

Qwen3.8 Flash is newer than the 32B and answers differently on hard prompts; the alias gets you running, then run your evaluation set before you stop looking. If you want Groq's exact replacement instead, openai/gpt-oss-120b is served here too at $0.20 / $0.80. Prices and capabilities: Qwen3.8 Flash here.

Option 3 — stop pinning one provider and one id

Put the model id in configuration and let a 404 or 429 fall through to a second provider. A retirement then degrades to a log line instead of an outage, and you choose when to migrate. LiteLLM's router does this in a few lines; the same shape exists in most frameworks.

LiteLLM proxy: Groq first, fall back on 404/429
# LiteLLM proxy config: Groq first, Inference APIs when Groq answers 404 or 429
model_list:
  - model_name: chat
    litellm_params:
      model: groq/openai/gpt-oss-120b
      api_key: os.environ/GROQ_API_KEY
  - model_name: chat-fallback
    litellm_params:
      model: inferenceapis/openai/gpt-oss-120b
      api_key: os.environ/INFERENCEAPIS_API_KEY

router_settings:
  fallbacks: [{ "chat": ["chat-fallback"] }]

The deprecations tracker lists every Groq retirement with its date and replacement, so you can see the next one coming.

Frequently asked questions

Is qwen/qwen3.8-27b on Groq the same model?

No. It is a newer, smaller Qwen release that Groq added to its list after retiring the 32B. Treat it as a new model and test it; the id is also different, so nothing resolves to it automatically.

How do I turn thinking off on Qwen3.8 Flash here?

Send reasoning_effort: "none" (reasoning: {enabled: false} and thinking: {type: "disabled"} are accepted too). The response then has no reasoning_content field and the answer comes back in one piece.

Why did the context window change?

Groq served Qwen 3 32B with a 131K window. Qwen3.8 Flash here takes up to 1,000,000 tokens in one request, and GPT-OSS 120B on Groq takes 131K. Check any code that trims history to a fixed size.

Where Inference APIs fits

Option 2 above keeps the old id working with a base-URL change: Inference APIs accepts qwen/qwen3-32b and serves it with Qwen3.8 Flash at $0.15 / $0.50 per million tokens, billed per request with no daily cap.

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