Model & Routing

The first group of Release settings: which model answers, what happens when it doesn’t, which languages it advertises, and how its traffic is routed.

Setting	Field	In the SDK?
Assistant model	assistant	via updateInWorkspace
Fallback chain	fallbackAssistants	✅
Languages	supportedLanguages	✅
AI Gateway	aiGateway	✅
Context cache	contextCache	✅

The assistant model

assistant is a { id, finetune } reference to a model from the catalog — over a hundred base and fine-tuned options across providers. Browse them in Model Catalog; training your own is Fine-tuning.

Fallback assistants

An ordered chain the Release falls through to when the primary model returns a retryable failure. First success wins.

await divinci.releases.updateInWorkspace(workspaceId, releaseId, {
  fallbackAssistants: [
    { id: "@cf/meta/llama-4-scout", finetune: false },
    { id: "gemini-3-5-flash", finetune: false },
  ],
});

Retryable means a provider 429, a 5xx, a capacity or quota error, or a timeout — the transient class. A refusal, a malformed prompt, or a content filter is not retryable and does not fail over: those would fail identically on the next model, so retrying would only cost more.

The tool loop has its own separate chain

fallbackAssistants covers the answering model. If the Release runs the skill/tool loop, its tool-calling model has an independent chain (skillConfig.fallbackToolCallingAssistants) and does not inherit this one. A Release whose assistant fails over cleanly can still lose tool use entirely when the tool model is the one having an outage — which presents as “the assistant stopped using its tools” rather than as an error. See Assistant Tools.

Languages

Which languages the Release advertises and will answer in. Unset, the set is derived from the model — the platform knows roughly what each family handles:

Tier	Size	Applies to
All 36	full platform set	Gemini family
Major (~19)	major world languages	Large open models — Llama 4, Llama 3.3 70B, Qwen3 30B, GLM-4, Kimi K2, Mistral Small, SEA-Lion…
Core (~9)	en es fr de pt it zh-hans ja ko	Small models (≤~20B) and anything unrecognised

Matching is a substring test on the lowercased model id, so it tolerates provider prefixes and version suffixes. Lock the set explicitly with custom:

await divinci.releases.updateInWorkspace(workspaceId, releaseId, {
  supportedLanguages: { mode: "custom", codes: ["en", "es", "fr"] },
});

Codes are BCP-47 (es, zh-hans). The resolved set drives two things: the server-side response-language allowlist — a language request outside the set is ignored — and the language switcher your client renders.

`en` is always included, and an unknown model gets the small set

A custom list is filtered to valid platform codes and always force-includes en, so a Release can never end up with no answerable language. Separately: a model the rules do not recognise falls back to the ~9-language core set, not to all 36. A brand-new large model may therefore advertise fewer languages than it actually speaks until a rule is added — set custom if that matters to you.

AI Gateway

Routes model traffic through a Cloudflare AI Gateway for analytics, rate limiting, and caching. It does not change how the model answers — same model, same prompt, same output; only the path differs.

await divinci.releases.updateInWorkspace(workspaceId, releaseId, {
  aiGateway: { enabled: true, useDefaultGateway: true },
});

Field	Default	Purpose
enabled	false	Master switch
useWhiteLabelDefaults	—	Inherit the workspace’s gateway settings — the “use workspace defaults” checkbox
useDefaultGateway	true	Use Divinci’s gateway (divinci-ai-gateway)
gatewayId	divinci-ai-gateway	Your own gateway id when useDefaultGateway is false
customAccountId	—	Your Cloudflare account id, for a BYO gateway
cacheTtlSeconds	3600	Sent as cf-aig-cache-ttl; bounds how stale a gateway cache hit can be

Leave useWhiteLabelDefaults on unless this Release specifically needs different routing from the rest of the workspace — that inheritance is what keeps gateway configuration in one place.

Gateway caching is a separate cache from context caching

cacheTtlSeconds caches whole responses at Cloudflare’s edge: an identical request inside the window gets a stored answer. Context caching instead discounts the repeated prefix of a prompt at the provider. They stack, they are configured independently, and only the gateway one can return a response the model never generated — so keep its TTL short on anything personalised.

Context cache

Cost control, not behaviour: implicit caching is free and on by default, and for most Releases needs no attention. Explicit mode adds a TTL you control and can cut input-token cost substantially on repeated prompts — worth reaching for once you have real volume, not before. Full detail in Caching & Performance.

Gotchas

• A fallback chain is not a failover for bad output. It fires on transport and capacity errors only. A model that answers badly answers badly.
• Every fallback entry should be a model you would be happy to serve. There is no quality gate between them — position in the array is the whole policy.
• Languages are advertised, not enforced on the model. The allowlist governs what your client offers and what the server accepts as a language request; it cannot stop a model replying in something else if asked to in the prompt.
• The gateway is observability, not a safety net. Enabling it does not add retries or failover — that is fallbackAssistants.

Related

• Model Catalog — the models available
• Fine-tuning & Training Data — training your own
• Caching & Performance — context caching in depth
• Releases — the object all of this lives on