Anuma says it has routed more than 1.2 million model requests through Bifrost, the open-source AI gateway developed by Maxim AI, since February. The gateway serves as a common routing layer between Anuma and multiple external AI providers, handling workloads spanning chat, reasoning, vision and tool execution without requiring separate application-level integrations for each provider.
According to a technical post published on the ZetaChain blog, Anuma reports more than 3 million total calls, including embeddings, and roughly 30 billion processed tokens, while its ZetaChain-backed private memory infrastructure is approaching 300,000 wallets. Those operating figures are project-reported metrics rather than independently audited usage statistics. ZetaChain has separately said Anuma has processed 30 billion private AI tokens.
Bifrost Handles Provider Routing and Failover
Bifrost operates as a self-hosted gateway between Anuma’s application layer and the model providers serving individual requests. Instead of connecting application code directly to every upstream provider, Anuma can route requests through a standardized interface, with models mapped to available backends and fallback providers available when the preferred route fails or becomes slow.
That architecture aligns with Bifrost’s documented capabilities. Maxim AI describes the open-source gateway as supporting more than 20 model providers through a single OpenAI-compatible API, with automatic failover, load balancing and unified provider management built into the routing layer. Bifrost can also track provider performance, latency, token consumption and costs.
Anuma says its implementation has maintained a failure rate below 0.1%, with routing logic able to move requests to another configured provider when the first option is unavailable. That redundancy is intended to separate the user-facing application from volatility at individual model providers, although the sub-0.1% figure comes from Anuma’s own operational reporting rather than a third-party benchmark.
Privacy Depends on Anuma’s Gateway Configuration
The routing layer also plays a role in Anuma’s privacy model. Anuma says its deployment limits gateway telemetry to operational information such as latency, cost and provider health rather than storing prompts and model responses as telemetry. Its broader platform architecture uses encrypted memory intended to remain portable across different AI models.
That distinction is deployment-specific. Bifrost itself supports much broader observability and, when configured accordingly, can record complete conversation inputs, model outputs and tool results. The absence of prompt and response logging therefore depends on how Anuma configures and operates Bifrost, not on an inherent limitation of the gateway software.
Anuma’s privacy policy similarly states that requests are transmitted to third-party model providers and that those providers operate under their own retention and data-handling policies. Anuma can minimize what its own infrastructure retains, but it cannot independently guarantee zero retention across every upstream model provider.
The project is now considering dynamic provider ranking based on live performance data and potentially opening parts of the routing infrastructure to external developers. The next material test is whether Anuma can extend the same routing and privacy controls beyond its own application while maintaining the reliability it reports at its current scale.
