Skip to content
Divinci AI SDK

Model Catalog

Copy page

The Divinci platform exposes one unified model catalog across every surface — the Divinci App, the @divinci-ai/* SDKs, the public API, the Divinci MCP server, and the Divinci CLI. Pick a model once and it resolves the same way everywhere: by its stable catalog id.

How the catalog works

Section titled “How the catalog works”
  • Catalog ids are stable. A model’s id (e.g. gemini-3.5-flash-lite, @cf/meta/llama-4-scout-17b-16e-instruct) is the key you pass to the SDK, API, MCP, and CLI. Persisted configs keep resolving even when a vendor renames the underlying runtime string.
  • Hosting differs by provider. Google-authored Gemini/Gemma models run on Google Vertex AI (Divinci pays — no key needed on the free tier). Open-weight models (@cf/...) run on Cloudflare Workers AI. You can also bring your own endpoint (byo:custom, local:...).
  • Free tier. Google-authored and Cloudflare-hosted models are free without a Bring-Your-Own-Key (BYOK). Everything else (OpenAI, Anthropic, Mistral, Moonshot/Kimi, …) is BYOK-only on the free plan but selectable once you attach a key. The SDK/CLI surface the same policy.
  • The list is dynamic. New releases (like the July 2026 gemini-3.5-flash-lite and gemini-3.6-flash Vertex launches) land in the server-driven catalog without a client redeploy. The table above is a curated snapshot — always confirm a specific id is available in your plan inside the App, SDK, API, MCP, or CLI.

Using a model id

Section titled “Using a model id”

The same id works identically on every surface:

import { divinci } from "@divinci-ai/sdk";


// App / SDK / API / MCP / CLI all accept the catalog id
const release = await divinci.releases.create(workspaceId, {
  name: "my-release",
  assistant: { id: "gemini-3.5-flash-lite", finetune: false },
});


// Or, in the Divinci CLI:
//   divinci release create --name my-release \
//     --assistant-id gemini-3.5-flash-lite

Search via query parameter

Section titled “Search via query parameter”

The table above filters live as you type. You can also deep-link a filtered view with the q query parameter — handy for docs, runbooks, and shareable links:

  • https://sdk.divinci.ai/guides/models?q=gemini → all Gemini models
  • https://sdk.divinci.ai/guides/models?q=vertex → Vertex-hosted models
  • https://sdk.divinci.ai/guides/models?q=llama → Llama / Meta models

The q value is synced to the URL (via history.replaceState), so the filtered state is shareable and the browser back button works.

Deprecations & migrations

Section titled “Deprecations & migrations”

When a vendor announces a model shutdown, Divinci captures it in the server-side model-deprecation watch (resources/server-resources/.../model-deprecations.ts) and migrates runtime usages to the replacement before the cutover date, so default invocations never 404. If you have persisted configs referencing a retired id, the stable-catalog-id pattern keeps them resolving to the GA replacement.