FAQ¶
Getting started¶
What does preview mean?¶
Verda Inference is live and ready for real workloads. During the preview, we're defining SLAs and SLOs for uptime and performance, usage limits apply, and details may change before general availability.
Which key do I use for inference requests?¶
Use an inference API key, sent as Authorization: Bearer <your key> to https://inference.verda.com. Cloud API credentials are for the Verda API at https://api.verda.com, such as managing models.
What do I send as model?¶
The model ID from the model's page, for example kimi-k3-1-pro. See Model IDs.
I get 403 subscription_required¶
The model isn't enabled for your project. Click Enable on the model's page. See Enable a model.
How much balance do I need?¶
At least $1 in your project balance to enable a model. The minimum top-up is $20.
I get 402, or the Console says inference is paused¶
Your project balance is too low. Top up, and inference continues. See Topping up in the console.
I get 429¶
You reached a rate limit, or the model is busy. Wait for the Retry-After time, then retry. See Handling 429 errors.
My streamed responses have no usage¶
Add "stream_options": {"include_usage": true} to the request. The last chunk then carries usage. See Streaming.
My request times out¶
A request without streaming times out after 60 seconds. A streamed response can run for up to 20 minutes, so stream long generations. See Timeouts.
Am I charged for failed requests?¶
No. Requests that fail or are interrupted, including timeouts, client disconnects and streams that end early, aren't charged.
How are reasoning tokens billed?¶
As output tokens. See Reasoning.
Why is content empty?¶
The model used all of max_tokens while reasoning, before it could answer, so finish_reason is length. Raise max_tokens, or set a lower reasoning_effort. See Reasoning.
Data and security¶
Do you use my prompts to train models?¶
No. Verda doesn't use your inputs or outputs to train any AI model. See the AI Terms of Service in Schedule 5 of our Terms of Service.
Do you store my prompts and responses?¶
No. Request context is held briefly in an in-memory inference cache and is removed within 2 hours. We keep only the operational metadata needed to run the service and bill you.
Is any data shared with the model provider?¶
No. Verda hosts the models on its own infrastructure. No request data is sent to the model developer or any other third party.
Can other customers see my cached context?¶
No. The cache is a shared in-memory performance layer. It's never exposed through the API, and no customer can read another customer's cached content.
Do you use subprocessors for inference?¶
No. Inference runs without subprocessors. The list of subprocessors for other Verda services is in the Trust Center.
Which certifications cover Verda Inference?¶
Verda is SOC 2 Type II, C5 and ISO 27001, 27017, 27018 and 27701 certified, and GDPR compliant. See the Trust Center for reports and details, and the Terms of Service for the terms that apply.
Models and usage¶
Where do the models run?¶
On Verda GPUs in our datacenters in Europe. Each model's page lists the country.
Can I get higher rate limits?¶
Yes. Contact support with the project, the model and the limits you need.
Can I use embeddings, fine-tuning or batch requests?¶
Not at the moment. Verda Inference serves Chat Completions and Responses. See OpenAI compatibility.
Can I use my own or a fine-tuned model?¶
Not on Verda Inference. Deploy it on serverless containers, or run your own stack on GPU instances.