Cloud Fallback
Live72 cloud models across nine providers, behind the same aliases as the on-prem fleet. Reached only when on-prem cannot serve the request.
Cloud fallback exists so that saturation or a node failure degrades latency rather than availability. 72 cloud models across nine providers sit behind the same tier aliases as the on-prem fleet, so a fallback is invisible to your code — the model ID you sent does not change.
Three on-prem models plus 72 cloud models is 75 live models on the gateway. Alias IDs in the catalog outnumber that, because several aliases resolve to the same pool — Riven Ultra Router explains the mapping.
The nine providers#
| Provider | Typically used for |
|---|---|
| OpenAI | General reasoning and broad tool-use compatibility |
| Anthropic | Long-context work and careful instruction following |
| Multimodal input and very long context windows | |
| xAI | Fast general chat |
| Mistral | Efficient open-weight lanes and European hosting |
| Perplexity | Search-grounded answers, paired with Riven Grounding |
| DeepSeek | Reasoning and code at low cost per token |
| Cohere | Embeddings and reranking |
| Groq | Very low-latency serving for short completions |
When fallback triggers#
- The on-prem pool for the resolved tier has no healthy replica with a free slot.
- The on-prem attempt fails or exceeds its deadline.
- The request needs a capability no on-prem model provides — a modality or a context length beyond the fleet.
- An organisation policy explicitly prefers a named provider for a given workload.
If none of these hold, the request stays on-prem. Fallback is a reaction to conditions, not a load-balancing default.
Knowing what served you#
Every response carries the upstream that produced it, so attribution never requires guessing.
| `x-riven-upstream` | Meaning |
|---|---|
onprem-qwen36-a | Served by the on-prem primary — no third party involved |
onprem-nemotron75-b | Served by the on-prem heavy-reasoning pool |
onprem-bonsai27-* | Served by the ternary efficiency pool |
| A provider name | Fell back to cloud — the provider is named explicitly |
Cost and quota#
Cloud tokens are metered through Riven Pricing Brain using the same tier multipliers documented in Models & usage. A fallback does not change the multiplier you are quoted — the tier you requested is the tier you are billed at.
Opting out#
Organisations that cannot send data to a third party at all can pin routing to on-prem only. Requests then queue or fail rather than falling back, which is the correct trade for regulated workloads. Configure it per organisation — see Riven Business Suite for which tiers include the policy.
Provider credentials are held in Riven Keymaster and are never exposed to clients or included in model context.