Z-ai/glm5 on Hosted Integrate appears unhealthy: repeated direct requests stall with zero bytes returned

AI & Data Science NVIDIA NIM Models
,
1

I can reproduce transport-level stalls directly against NVIDIA Hosted Integrate for:

This was reproduced directly against NVIDIA’s endpoint, not only through a local proxy.

Problem

For simple requests, the route stalls at the transport layer:

  • HTTP 200 is never returned
  • zero bytes are ever received
  • the caller hangs indefinitely until timeout (tested up to 150 seconds)
  • this is not a malformed response; the model route simply never produces a usable body

This is distinct from content-quality or json-formatting issues seen in related threads.

Minimal failing repro

curl -sS https://integrate.api.nvidia.com/v1/chat/completions \
-H "Authorization: Bearer $NVIDIA_API_KEY" \
-H "Content-Type: application/json" \
--max-time 150 \
-d '{ "model": "z-ai/glm5", "messages": [{"role": "user", "content": "Return exactly: OK"}], "max_tokens": 32, "stream": false }'

Observed failing result

In repeated independent requests across separate runs:

  • curl times out after the specified --max-time duration
  • no response is ever returned
  • zero bytes are received
  • the route stalls at transport level, not application level

Control

In the same general environment, repeated requests to a baseline model such as meta/llama-3.1-8b-instruct returned HTTP 200 OK with exact “OK” content.

So this does not look like a total platform outage. It looks like a z-ai/glm5 route-health problem on Hosted Integrate.

Related concerns

We have also seen other reports involving hosted glm5 instability, including:

  • malformed tool-call JSON
  • 500 “EngineCore encountered an issue” errors

Even when the route is listed in /v1/models, it is not behaving like a healthy chat-completions route.

Temporary mitigations on our side

  • strip reasoning_effort
  • strip response_format
  • fail closed on first-response stalls instead of letting callers hang indefinitely
  • avoid trusting streaming as a sign that the route is healthy
  • on our proxy, this route consistently returns 504 Gateway Timeout under load

These mitigations help clients fail more cleanly, but they cannot fix a route that never produces a first byte.

Request

  • please check the hosted glm5 deployment / scheduling / routing health on Integrate
  • if glm5 is degraded, please stop advertising it as healthy in /v1/models until it passes real chat-completions checks
  • if glm5 currently requires a very specific request profile on this endpoint, please document that explicitly

Minimal repro snippet

curl -sS https://integrate.api.nvidia.com/v1/chat/completions \
-H "Authorization: Bearer $NVIDIA_API_KEY" \
-H "Content-Type: application/json" \
--max-time 150 \
-d '{ "model":"z-ai/glm5", "messages":[{"role":"user","content":"Return exactly: OK"}], "max_tokens":32, "stream":false }'