Tencent's open-source 770B MoE model with 1M-token context arrives on Vercel AI Gateway

Tencent's Hy4 Preview lands on Vercel AI Gateway

Vercel has announced that Hy4 Preview, an open-source model from Tencent, is now reachable through its AI Gateway service. According to Vercel's changelog post, the model uses a Mixture-of-Experts (MoE) architecture with 770 billion parameters in total, of which 49 billion are active for any given token, and it accepts a context window of one million tokens.

What the model is built for

Vercel positions Hy4 Preview at workloads that reward both scale and long context. The company lists sustained coding sessions, working through lengthy documents, game development and scientific reasoning as the intended use cases. The sparse design — a large total parameter count with only a fraction engaged per token — is a common strategy for keeping serving costs and latency down while preserving the capacity of a much denser model. The announcement includes no benchmark figures, so developers will need to weigh the preview against other models themselves.

How developers can call it

The model is exposed under the identifier tencent/hy4-preview through the AI SDK. Vercel's example shows a simple streaming call: set the model string, supply a prompt — its sample asks the model to add pagination to an API endpoint — and consume the stream.

There is also a path for tooling that cannot call the SDK directly. Running vercel ai-gateway coding-agents setup, per Vercel's coding agents guide, connects tools such as Claude Code, Codex, OpenCode, Cursor and Pi to the gateway, after which Hy4 Preview can be selected from inside the agent. Vercel additionally offers a model playground for trying the model before wiring it into an application.

What AI Gateway contributes

AI Gateway is Vercel's routing layer for language models: a single API surface in front of many providers, plus usage and cost tracking, retries and failover, and performance optimisations that Vercel says are aimed at exceeding the uptime of any individual provider. The service also ships per-key budgets, routing rules, custom reporting and Zero Data Retention support for teams with strict data-handling requirements.

On pricing, Vercel says the gateway passes provider rates through without a markup and charges no platform fee on inference, including when customers bring their own provider keys under its BYOK option.

Why it matters

Two things make this release notable. First, an open-source model of this scale — 770B parameters paired with a one-million-token context window — becoming available through a mainstream deployment platform lowers the barrier for teams that want large-model capacity without negotiating direct relationships with a model provider or standing up their own serving infrastructure.

Second, the coding-agent integration matters beyond this single model. Tools like Claude Code and Cursor were originally built around specific vendors; gateway-based access turns them into generic clients for anything the gateway serves, and a newly released model gains a ready-made distribution channel on day one. As the name indicates, Hy4 Preview is still a preview, so teams adopting it should treat its behaviour and availability as subject to change.