Hy4 Preview now available on AI Gateway: Hy4 Preview from Tencent is

What happened

Vercel's AI Gateway now supports Hy4 Preview, the open-source Mixture-of-Experts model from Tencent, giving developers direct access to a 770-billion-parameter system through a single unified routing layer. The addition expands the gateway's model catalog without requiring teams to manage separate API credentials or provider integrations.

This article walks through exactly what Hy4 Preview is, how its architecture differs from dense transformer models, what the 1-million-token context window means in practice, how to wire it into an existing AI SDK project, and where the rough edges are likely to appear during early adoption. It is aimed at full-stack and backend engineers already building on Vercel's infrastructure who want to evaluate whether this model fits their workloads.

Tencent's Hy4 Preview is a Mixture-of-Experts model with 770 billion total parameters, now routed through Vercel's AI Gateway. Mixture-of-Experts architectures activate only a subset of parameters per inference pass, so the effective compute cost per token is lower than a dense model of equivalent nominal size. Hy4 Preview is open-source, meaning the weights are publicly available for inspection and fine-tuning, and it carries a context window of 1 million tokens. Vercel is hosting access to it through AI Gateway under the identifier tencent/hy4-preview, making it reachable from any codebase that already talks to the gateway.

How it works

Tencent positions Hy4 Preview as targeting four domains: long-horizon coding tasks, document analysis, game development, and scientific reasoning. The 1-million-token context window is the practical enabler for most of those use cases, since long codebases, legal documents, game-state histories, and research papers all routinely exceed the 128k or 200k limits common on other hosted models. The preview designation signals the model is still being evaluated and tuned, which is worth keeping in mind before committing to it in a production critical path.

The primary change is model availability, not API surface. If you are already calling Vercel's AI Gateway through the AI SDK, the integration point is a single model string: tencent/hy4-preview. No new SDK version, no separate Tencent API key, and no additional configuration block is required beyond swapping that identifier in wherever you currently specify a model. The gateway handles credential forwarding, request routing, and response normalization.

Why it matters

The more significant shift is architectural opportunity. A 1-million-token context window changes what is feasible to send in a single request. Teams working on code review tools, contract analysis pipelines, or simulation-heavy game backends can now pass far more context without chunking, summarizing, or building retrieval systems to compensate for a shorter window. That removes a class of engineering complexity, though it also pushes token costs higher per request, so the economics need to be evaluated per workload rather than assumed to be favorable.

To use Hy4 Preview, set the model field to tencent/hy4-preview in your AI SDK configuration. If you are already routing requests through AI Gateway, no package update is required on the gateway side. Verify your AI SDK dependency is current enough to pass arbitrary model identifiers through to the gateway without validation errors, since older versions sometimes maintained a hard allowlist of known model names.

If you are starting fresh, the standard path is to initialize a project using Vercel's AI SDK, configure the AI Gateway endpoint as your provider base URL, and then set model to tencent/hy4-preview in the model instantiation call. Test with a short prompt first to confirm routing is working before sending large context payloads, since a misconfigured endpoint will fail immediately on a small request and save you from burning tokens on a malformed large one.

Who is affected

Preview-stage models carry implicit instability. Tencent may update weights, change the model's behavior, or alter rate limits without the same notice period that a stable release would receive. Any evaluation you run today should be repeated if you observe behavioral drift in the coming weeks, and production pipelines should have a fallback model configured until the model exits preview status.

The 1-million-token context window is the feature most likely to create unexpected cost spikes. Teams migrating from models with shorter windows may discover that existing prompt construction logic, which previously relied on truncation or retrieval to stay within limits, now passes everything through unchecked. Auditing your context assembly logic before enabling Hy4 Preview in any automated pipeline will prevent runaway token usage. Also confirm that your Vercel plan and AI Gateway tier support the throughput you expect, since large-context requests occupy more infrastructure resources per call.

What to watch next

The most important signal to track is when Tencent moves Hy4 Preview to a stable release designation. That transition will indicate that the weights are frozen and the interface is considered production-ready, which is the appropriate moment to evaluate it for customer-facing features. In parallel, watch for any published benchmark results specifically covering the four stated domains: long-horizon coding, document analysis, game development, and scientific reasoning, since Tencent's own characterization is the only public framing available at this stage.

On the Vercel side, monitor whether AI Gateway adds usage analytics or per-model cost dashboards that surface Hy4 Preview's token consumption distinctly. Large-context models are difficult to reason about financially without per-request visibility, and tooling that makes that visible will lower the risk of adopting the model more broadly across a team's projects.