Error 504: what to do when qwen3-8b fails

A 504 means the gateway timed out waiting for the upstream model: the request was sent, but the upstream did not return within the gateway timeout.

504 Gateway Timeout means the gateway gave up waiting for the upstream response. The request was forwarded, but no result came back within the gateway limit. Shortening the output, switching to streaming, or lowering max_tokens usually fixes it.

qwen3-8b is served by Alibaba. Everything on this page — triggers, fixes and measured data — is compiled from the real runtime behaviour of this model at the gateway layer.

At this gateway, the most common trigger is: The upstream took longer than the gateway timeout allows. The recommended first action is: Switch to streaming output.

Common causes

  • The upstream took longer than the gateway timeout allows
  • A long context pushed inference time up
  • An upstream node is stuck
  • A non-streaming request with a very long output

How to fix

  • Switch to streaming output
  • Shorten the context and lower max_tokens
  • Retry after a short wait
  • Use a model with faster inference

Retry with exponential backoff

The snippet below retries when qwen3-8b returns 504, up to 5 attempts, with an increasing wait plus random jitter so concurrent calls do not retry in lockstep. Read the base URL and API key from environment variables — never hardcode them.

import os, time, random
import requests

BASE  = os.getenv("OPENAI_BASE_URL")   # e.g. https://<your-gateway>/v1
KEY   = os.getenv("OPENAI_API_KEY")
MODEL = 'qwen3-8b'


def chat(messages, retries=5):
    """Retry with exponential backoff + jitter."""
    for i in range(retries):
        try:
            r = requests.post(
                BASE + "/chat/completions",
                headers={"Authorization": "Bearer " + KEY},
                json={"model": MODEL, "messages": messages, "stream": True},
                timeout=60,
            )
            if r.status_code == 429 or r.status_code >= 500:
                time.sleep(min(2 ** i + random.uniform(0, 1), 30))
                continue
            r.raise_for_status()
            return r.json()
        except requests.exceptions.Timeout:
            time.sleep(min(2 ** i + random.uniform(0, 1), 30))
    raise RuntimeError("gave up after " + str(retries) + " retries")


print(chat([{"role": "user", "content": "hello"}]))

Key facts for this model

API endpointhttps://api.airai.cc/v1
OpenAI-compatibleOpenAI-compatible
VendorAlibaba
Context131.1K
CapabilitiesReasoning, Tools, Open Weights
API formatsopenai, openai-response, openai-response-compact, anthropic, gemini, openai-alpha-search
Billing formulap * 0.18 + c * 2.1) : tier("standard", p * 0.18 + c * 0.7

FAQ

How much concurrency is safe?

Lower the concurrency first — most throughput complaints disappear once you do. This model has a 131.1K context window and comes from Alibaba. A 131.1K context means long inputs add noticeably to first-token latency. For long outputs, raise the timeout to 60 seconds or more. Use batching or a queue to smooth peaks — steadier than raising concurrency on the fly.

Does switching to a smaller model reduce 504?

Building a fallback into the architecture is more reliable than patching errors one by one. This model from Alibaba has several upstream nodes the gateway can switch between. Keep a lighter fallback model ready so the main flow never breaks. Put the model name in config, so switching upstreams needs no code change.

What information should I give support when reporting an issue?

Group the errors by time and node first; the pattern is usually obvious once you do. If it only happens in production, it is usually an environment difference, not the model. This model is served by Alibaba, so upstream status follows the vendor’s own announcements. Log the request ID on every failure — it beats the status code when debugging. Reproduce it once in a staging environment with the same request body.

Will I be charged when qwen3-8b returns 504?

Only the output already produced is billed; the failed part is not. Billing follows p * 0.18 + c * 2.1) : tier("standard", p * 0.18 + c * 0.7, so no output means no charge. Input price is about $0.18 per million tokens. When estimating cost from p * 0.18 + c * 2.1) : tier("standard", p * 0.18 + c * 0.7, include the retry budget.

Other errors on this model

Other models with the same error

Back to the model page

Data updated: 2026-10-10 12:10

Technical SupportLive Support
Back to Top