I’ve noticed a consistent latency issue on all my calls on the first user question. Specifically, after the user finishes speaking, there is a noticeable delay (around 3 seconds in the example) before the model request becomes active. Subsequent responses are much faster, which suggests this might be related to the initial session setup or model initialization. From the logs (see attached image), the delay occurs after endpointing detection and KB retrieval but before the first model request starts. My assumption is that the initial model load could be the cause. I’d appreciate your guidance on why this happens. It doesn’t seem to be related to KB retrieval, and I’d also like to know if there are best practices to avoid this, perhaps by adjusting settings or using a different model ? Additionally, I was wondering if sending a placeholder tool request that triggers at the beginning of the call could help "make the model ready". Is this a viable approach? Example call IDs: - 6e6f1c1f-814b-48a7-8222-97fff942963e - 89d6d8fd-d989-406b-9944-309c27105eed Thank you in advance for your help! https://cdn.discordapp.com/attachments/1398342221056708638/1398342221450842132/Screenshot_2025-07-25_at_8.56.10_a.m..png?ex=6885032d&is=6883b1ad&hm=d18733883609a5103098292a1fc320fdaa02692cc3e9c9b3daed790ee6c4cbb1&

The first-response latency you see is most likely a model cold-start/initialization delay. While you can shave off some time by tuning delay settings and picking optimized models, there's currently no built-in warm-up mechanism in Vapi. A placeholder tool request might help marginally, but better overall performance comes from reducing buffer settings or using cache-friendly prompts.

Hey! To help track down this issue, could you share: \- The call ID \- When exactly this happened (the timestamp) \- What response you expected to get \- What response you actually got instead This would really help us figure out what went wrong!

call id: 6e6f1c1f-814b-48a7-8222-97fff942963e timestamp: July 24, response I expected: not a response of 3 seconds, response I got: a response of 3 seconds.

Hey, thanks for sharing your details. Our team will look into it and get back to you soon with an update.

Hi, thank you, any updates on this ?

any updates? I'm also seeing this cold start delays

any updates facing same issue

Hi alep, If you're experiencing significant latency, you can consider adjusting the voice pipeline configuration to reduce response time. Here are some key actions: 1. **Optimize Start Speaking Plan**: - Use Smart Endpointing for English conversations to predict when users have finished speaking. [See configuration](https://docs.vapi.ai/customization/voice-pipeline-configuration). - Use the aggressive wait function for quicker responses in real-time interactions:

"waitFunction": "2000 / (1 + exp(-10 * (x - 0.5)))"

. 2. **Adjust Stop Speaking Plan**: - For sensitive environments, set

voiceSeconds

to a balanced level like

0.2

and use

numWords = 0

for VAD-based detection to improve responsiveness. 3. **Backoff and Wait Seconds**: - Ensure

backoffSeconds

and

waitSeconds

are set to create an optimal pause that suits your interaction speed. For fast-paced interactions, shorten backoff seconds to

0.5

. 4. **Language Considerations**: - If using non-English languages, configure transcription-based endpointing, adjusting wait times around punctuation and numbers. 5. **Flush Syntax**: - Implementing flush syntax can ensure acknowledgments are immediate during processing. Ensure

chunkPlan

is enabled when using flush syntax. [Learn more about flush syntax usage](https://docs.vapi.ai/assistants/flush-syntax). By applying these configurations, you can effectively reduce latency during conversations. Let me know if you need further assistance with specific settings.