Hi Nvidia Team,
The Nvidia / GLM-5.2 model endpoint is consistently failing via a Vercel AI SDK integration in my IDE. Other models work perfectly, isolating the issue to this specific backend stream.
Error: AI SDK stream stalled: no event for 60000ms (60s)
Behavior: The connection opens, but the server goes silent and fails to stream tokens, even on short prompts.Could you please verify if the GLM-5.2 API endpoint is currently experiencing routing bugs or server-side drops?
Thank you.
I’m not sure what you are talking about as this is the Workbench forum.
However, you can try help@build.nvidia.com for these kinds of issues.
The 60s stall usually means the server accepted the request but never produced a first token, so the AI SDK client times out. For GLM-5.2 specifically, this is often one of three things:
1. Public API Endpoints entitlement. Newer frontier models on personal/free orgs can return 404 or silently hang if the org lacks the Public API Endpoints permission. Check build.nvidia.com org settings.
2. GLM-5.2 context/prompt handling. Some providers silently error on system prompts or certain JSON schema modes. Try a minimal user-only message with no system prompt and low max_tokens.
3. Streaming compatibility. The Vercel AI SDK expects a particular SSE format. If the endpoint returns a non-streaming JSON response under some error condition, the SDK sees no events and hits the stall timeout.
Quick isolation:
- Test the same model with a raw curl to integrate.api.nvidia.com/v1/chat/completions using stream=true and a short timeout. If curl also hangs, it is server-side or account-scope.
- If curl works but the SDK fails, add experimental_stream to the SDK call or compare response headers.
For account/entitlement issues, email help@build.nvidia.com with the org ID and endpoint. For SDK-specific behavior, the build.nvidia.com forum or the AI SDK repo are better fits than Workbench.