How can I make Grok 4 Fast work in AI Foundry? Playground gives Error : no healthy upstream only for Grok 4 Fast

Question

How can I make Grok 4 Fast work in AI Foundry? Playground gives Error : no healthy upstream only for Grok 4 Fast

Wessel Vinke 5

I deployed Grok 4 Fast (reasoning and non-reasoning version) in multiple locations (Sweden, West Europe, and US). However, in the chat playground I always get:

Error

: no healthy upstream Please check this guide to understand why this error code might have been returned https://docs.microsoft.com/en-us/azure/machine-learning/how-to-troubleshoot-online-endpoints#http-status-codes

This does not happen for other models I've tried.

How can I make sure that this works, so that Grok 4 Fast can be implemented via API?

Ashwani kumarsahu 0 Reputation points

2025-11-03T15:45:44.2033333+00:00
Hello Wessel,

I faced a similar issue while deploying Grok 4 Fast. The error "no healthy upstream" usually occurs due to region mismatch or inactive endpoints.

Here are a few things to check:

Confirm that the region supports Grok 4 Fast.

Validate your endpoint and configuration file.

Check Azure quota limits and model health status.

Try redeploying in a supported region or wait a few hours and retry. Hope this helps!

Regards,

Ashwani
Anonymous

2025-11-04T09:16:23.61+00:00

Hello Wessel Vinke,

Good day.

There is an outage declared for Grok 4 Fast, and the engineers are working on it.

Apologies for the inconvenience caused. Will keep you posted on the update.
Wessel Vinke 5 Reputation points

2025-11-04T11:33:47.9233333+00:00

It seems that there was an outage, and the outage has now been resolved Update: now I'm getting an error again with grok 4 fast reasoning in specific (503)
Insaf Elinkichari 0 Reputation points

2025-11-04T13:47:43.21+00:00

Hello @Wessel Vinke
The error "no healthy upstream" indicates that the deployment endpoint is created, but no model instance became ready behind the load balancer. This usually happens in AI/ML Foundry when the model container fails during startup.

For Grok 4 Fast specifically (reasoning and non-reasoning), there are a few checks that resolved it for us:

Grok 4 Fast requires GPU SKUs that support large memory. If the deployment auto-selects a GPU that is not supported (or quota is exhausted), the container never becomes healthy.

Fix:

Open AI Hub → Projects → Deployments

Edit deployment → Compute

Select a supported GPU SKU manually (for example: Standard_NC24ads_A100_v4 or Standard_ND96asr_v4 depending on the region)

Also verify GPU quota: Azure Portal → Subscriptions → Quotas → GPU
Anonymous

2025-11-06T04:49:56.7533333+00:00

Hello Wessel Vinke,

Updating you on the status grok 4 fast reasoning, engineers are still working on resolving the issue.

Thank you for being patient.
Anonymous

2025-11-10T01:25:58.6466667+00:00

Hello Wessel Vinke,

Try redeploying your model as both Grok 4 Fast reasoning and non-reasoning version) is working fine in Sweden central and East US 2.

Thank you for being patient.
Anonymous

2025-11-14T04:08:40.62+00:00

Hello Wessel Vinke,

Update on grok 4 models reasoning & non-reasoning:

They are still having in-consistent behavior in different regions and still facing issues.

PG team is working on resolving the issue. Apologies for the impact it's causing on your work.

Thank you for patience.
Anonymous

2025-11-17T05:51:49.84+00:00

Hello Wessel Vinke,

Thank you for your patience.

Update on current status:
There is no confirmed root cause yet. The issue appears to be an intermittent engine crash resulting in occasional 503 / timeout responses.
Microsoft PG engineers are in sync with xAI team to determine next troubleshooting steps and continue investigation.

I will keep you updated as soon as the engineering team provides any new findings.
Anonymous

2026-02-04T13:28:19.0366667+00:00

Hi Wessel Vinke

We apologise for the delay;Microsoft PG engineers are in sync with xAI team the issue got Mitigated by the xAI container enabled disaggregated PD to optimize xAI container's throughput and reduce latency,getting more capacity to make the model more responsive under production traffic.please try to check from your side
Anonymous

2026-02-05T10:02:31.1733333+00:00

Hi Wessel Vinke

Did you get any chance to review the above response.

Thank you!

Your answer

Ashwani kumarsahu 0 Reputation points

2025-11-03T15:45:44.2033333+00:00

Hello Wessel,

I faced a similar issue while deploying Grok 4 Fast. The error "no healthy upstream" usually occurs due to region mismatch or inactive endpoints.

Here are a few things to check:

Confirm that the region supports Grok 4 Fast.

Validate your endpoint and configuration file.

Check Azure quota limits and model health status.

Try redeploying in a supported region or wait a few hours and retry. Hope this helps!

Regards,

Ashwani
Anonymous

2025-11-04T09:16:23.61+00:00

Hello Wessel Vinke,

Good day.

There is an outage declared for Grok 4 Fast, and the engineers are working on it.

Apologies for the inconvenience caused. Will keep you posted on the update.
Wessel Vinke 5 Reputation points

2025-11-04T11:33:47.9233333+00:00

It seems that there was an outage, and the outage has now been resolved Update: now I'm getting an error again with grok 4 fast reasoning in specific (503)
Insaf Elinkichari 0 Reputation points

2025-11-04T13:47:43.21+00:00

Hello @Wessel Vinke
The error "no healthy upstream" indicates that the deployment endpoint is created, but no model instance became ready behind the load balancer. This usually happens in AI/ML Foundry when the model container fails during startup.

For Grok 4 Fast specifically (reasoning and non-reasoning), there are a few checks that resolved it for us:

Grok 4 Fast requires GPU SKUs that support large memory. If the deployment auto-selects a GPU that is not supported (or quota is exhausted), the container never becomes healthy.

Fix:

Open AI Hub → Projects → Deployments

Edit deployment → Compute

Select a supported GPU SKU manually (for example: Standard_NC24ads_A100_v4 or Standard_ND96asr_v4 depending on the region)

Also verify GPU quota: Azure Portal → Subscriptions → Quotas → GPU
Anonymous

2025-11-06T04:49:56.7533333+00:00

Hello Wessel Vinke,

Updating you on the status grok 4 fast reasoning, engineers are still working on resolving the issue.

Thank you for being patient.
Anonymous

2025-11-10T01:25:58.6466667+00:00

Hello Wessel Vinke,

Try redeploying your model as both Grok 4 Fast reasoning and non-reasoning version) is working fine in Sweden central and East US 2.

Thank you for being patient.
Anonymous

2025-11-14T04:08:40.62+00:00

Hello Wessel Vinke,

Update on grok 4 models reasoning & non-reasoning:

They are still having in-consistent behavior in different regions and still facing issues.

PG team is working on resolving the issue. Apologies for the impact it's causing on your work.

Thank you for patience.
Anonymous

2025-11-17T05:51:49.84+00:00

Hello Wessel Vinke,

Thank you for your patience.

Update on current status:
There is no confirmed root cause yet. The issue appears to be an intermittent engine crash resulting in occasional 503 / timeout responses.
Microsoft PG engineers are in sync with xAI team to determine next troubleshooting steps and continue investigation.

I will keep you updated as soon as the engineering team provides any new findings.
Anonymous

2026-02-04T13:28:19.0366667+00:00

Hi Wessel Vinke

We apologise for the delay;Microsoft PG engineers are in sync with xAI team the issue got Mitigated by the xAI container enabled disaggregated PD to optimize xAI container's throughput and reduce latency,getting more capacity to make the model more responsive under production traffic.please try to check from your side
Anonymous

2026-02-05T10:02:31.1733333+00:00

Hi Wessel Vinke

Did you get any chance to review the above response.

Thank you!