Skip to main content
This page is the mental model for the dashboard. Deploy details live in Launch a model. How you pay is in Billing.

​
The three products

A vector store id is the knowledge base id. There is no second product.

​
What you can launch

A model is any supported pipeline, not only a chat LLM. The same launch wizard and worker settings apply to embeddings, vision, speech, classification, and the other Hub tasks we list. The pipeline_tag (or the artifact’s pipeline) chooses the API routes.

​
Launch, workers, and sleep

Each model has minimum and maximum workers.
  • Minimum 0 plus an instance timeout lets the last worker shut down when idle (Sleeping). The next request loads capacity again.
  • Minimum ≥ 1 keeps at least one worker loaded (Ready) for lower latency.
  • Scale-up threshold and cooldown control how quickly extra workers appear. See Autoscaling.
The launch wizard shows a live cost estimate before you confirm.

​
Pricing plans

Hourly (default): billed from GPU memory (VRAM) used by active workers. Precision, quantization, and LLM context length change VRAM. Per token (beta): available for some text-generation launches when minimum workers is 0 and the account flag allows it. Instance timeout is forced to 0. The model stays in a sleeping standby until traffic arrives. Exact rates are on Usage (Live pricing) and in the launch wizard. Do not hard-code dollar amounts from this page. Prepared artifacts also incur storage (GB-month). Agents can add tool usage (for example web search) and knowledge-base search charges shown in usage.

​
Credits

Credits map 1:1 to US dollars on the Billing page. They pay for compute, storage, knowledge, and tools. They do not expire and are not transferable. Add credits and automatic top-up keep workers from stopping on an empty balance. The account is in public beta; what you see in Billing is authoritative.

​
Model lifecycle

Track this on the model page. See Manage models.

​
Data store

On launch (and on agents) you can Store requests. When enabled, matching chat/completions/responses (and agent runs) appear on the Data tab. This is opt-in debugging data for your account, not training.

​
Infrastructure

Workloads run on shared, multi-tenant GPU capacity. Requests use HTTPS. Payloads are processed to complete the call and are not sold. Isolation, encryption, and retention are described in Data and the Security policy.

​
API shape

  • Base URL: https://api.synapsai.cloud (versioned routes under /v1)
  • Auth: Authorization: Bearer + API key
  • Python: pip install --upgrade synapsai-python (SynapsAI client)
  • OpenAI libraries: same routes for chat, embeddings, images, audio, vector stores — see Migrate from OpenAI and Integrations

Launch your first model

Account, key, and first chat completion.