Tavus incident
Conversations using Gemma-4 are experiencing delays
Tavus experienced a minor incident on September 14, 2026 affecting Conversational Video Interface, lasting 1h 4m. The incident has been resolved; the full update timeline is below.
Affected components
Update timeline
- investigating Sep 14, 2026, 05:58 PM UTC
We are aware of an issue currently impacting tavus conversations, we are investigating.
- investigating Sep 14, 2026, 06:03 PM UTC
It appears this issue is coming from one of our downstream providers, Crusoe. We're continuing to investigate full scope of conversations affected.
- investigating Sep 14, 2026, 06:14 PM UTC
Starting at 17:40 UTC, conversations using our default LLM began experiencing long delays and unresponsive replicas. A GPU provider is returning errors and our automatic failover is recovering most turns, but with significant added latency. We are routing traffic to a backup provider.
- identified Sep 14, 2026, 06:18 PM UTC
This is affecting gemma-4; we recommend our customers switch to a different LLM for immediate needs.
- identified Sep 14, 2026, 06:25 PM UTC
The issue is isolated to our default LLM (gemma-4), served by a GPU provider that began returning errors at 17:40 UTC. As a temporary mitigation we are routing gemma-4 traffic to an alternate model. Customers on gemma-4 may notice a change in response style while this is in effect; we will revert once the provider recovers.
- identified Sep 14, 2026, 06:42 PM UTC
We've confirmed the issue originates with our GPU provider, who has declared a networking incident on their side. We're testing mitigations to route gemma-4 traffic elsewhere. Conversations using other LLMs are unaffected. If you need stability now, you can point your persona at a different LLM.
- monitoring Sep 14, 2026, 06:55 PM UTC
Our upstream GPU provider has applied a mitigation and gemma-4 error rates have returned to normal. We have alternate routing paths staged and ready to enable if the issue recurs. We are monitoring before marking this resolved.
- resolved Sep 14, 2026, 07:02 PM UTC
This incident is now resolved.