Vercel Adds Kimi K3 With US Inference and Zero-Data Retention.

Vercel has added Moonshot AI’s Kimi K3 to AI Gateway, giving developers access to the open-weight model through US-based inference providers, automatic failover, and optional Zero-Data Retention agreements. A lower-latency serving option, Kimi K3 Fast, is available alongside the standard model.

The company described Kimi K3 as the world’s most powerful open-weight model in its LinkedIn announcement. More concretely, the release addresses three practical considerations for teams deploying AI applications: where inference runs, what happens to submitted data, and how applications keep operating when an individual model provider has problems.

According to Vercel’s official release, Kimi K3 and Kimi K3 Fast are available through US-based providers including Baseten and Fireworks. The models were added on July 27, 2026.

Developers can call Kimi K3 through the moonshotai/kimi-k3 model ID. AI Gateway then selects an underlying inference provider and can route requests to another provider if the first becomes unavailable.

That abstraction is a central part of the update. Instead of integrating separate provider APIs, handling different credentials, and writing custom fallback logic, teams can send requests through one gateway interface. Vercel says its multi-provider setup offers higher uptime and more available throughput than relying on a single provider.

The US-based serving options may also make the model practical for organizations with data residency policies or compliance requirements that limit where workloads can run. However, using US infrastructure does not by itself establish that an application meets a particular regulatory standard. Teams still need to assess their full data flow, storage configuration, logging, and downstream services.

Vercel is continuing to broaden AI Gateway beyond text models. The Kimi release follows its addition of Grok Imagine Image 2.0 to AI Gateway, while integrations such as AI Gateway and Sandbox support on Hermes Agent show how the gateway is becoming part of a wider agent deployment stack.

Kimi K3 Fast uses a faster serving path for applications where response time matters more than minimizing token costs. Vercel says the fast option carries a higher per-token price, although the supplied announcement does not include complete pricing figures.

Developers have two ways to request it:

• Use the speed option with the base moonshotai/kimi-k3 model. If the faster tier is unavailable, AI Gateway falls back to standard speed.
• Call moonshotai/kimi-k3-fast directly when the application should explicitly target the fast variant.

The first approach is likely to be more resilient for customer-facing products because it preserves a standard-speed fallback. The dedicated model ID gives developers more explicit control over model selection.

Zero-Data Retention, or ZDR, is supported for both versions. Vercel says it has signed ZDR agreements with the relevant providers and that customers can enable the protection for token traffic. In practical terms, this is designed to prevent model providers from retaining prompts and generated responses after processing, an important consideration when applications handle proprietary or sensitive information.

For small product teams, the update reduces the operational work involved in testing and deploying Kimi K3. Provider routing, fallback, US inference, and a faster serving tier can all be managed through the gateway rather than assembled as separate integrations. The remaining decisions are application-specific: whether ZDR is required, whether the latency improvement justifies the higher token cost, and whether the standard fallback behavior fits the product’s reliability requirements.

What is Vercel Kimi K3?

Kimi K3 is an open-weight Moonshot AI model now available through Vercel AI Gateway. Vercel provides access through US-based inference providers and handles provider selection and fallback behind a single model ID.

What is the difference between Kimi K3 and Kimi K3 Fast?

Kimi K3 Fast uses a lower-latency serving path but costs more per token than standard Kimi K3. Developers can request it using the speed option on the base model, with fallback to standard speed, or call moonshotai/kimi-k3-fast directly.

Does Kimi K3 support Zero-Data Retention on Vercel?

Yes. Vercel says ZDR is supported for both Kimi K3 and Kimi K3 Fast through agreements with its inference providers. Teams can enable it for token traffic, but should still review their broader application architecture and compliance obligations.