Tencent Hy4 is Tencent’s November 2026 multimodal model for coding, long-context work and agent workflows. Hosted access is real: Vercel added it to AI Gateway on November 3, and OpenRouter lists a separate route. The model’s current weights status requires more care than the phrase open source suggests.

This guide owns the current Hy4 integration decision. The earlier Hy4 Preview checkpoint and the Hy3 guide describe different releases. A provider listing for Hy4 does not prove that an older preview repository contains the same weights. We have not independently benchmarked Hy4 or verified a self-hosted deployment of its final checkpoint.

Current access and identity

Item	Verified route information
Vercel gateway identifier	tencent/hy4
OpenRouter identifier	tencent/hy4
Context	1,048,576 tokens on OpenRouter; 1M in Vercel’s announcement
Inputs	Text and image on the documented hosted product
Output	Text
Documented features	Reasoning, tool calling, structured outputs
Final weights	Tencent promises release; availability not independently confirmed in this check

Tencent’s November 4 announcement says weights and system cards will be released in the coming days. That is an announced release plan, not a verified download. Keep it separate from the Hy4 Preview repository.

This distinction matters for licensing, hardware and reproducibility. Do not attach a preview license or parameter count to the final release without checking the actual final artifacts. A model name selected in a gateway does not expose its underlying checkpoint revision in the same way as a pinned local deployment.

Calling the model through Vercel

The gateway announcement documents AI SDK use with the provider-qualified model name:

import { streamText } from 'ai';

const result = streamText({
  model: 'tencent/hy4',
  prompt: 'Outline a database migration in three brief steps.',
});

This is a documentation-derived example, not a paid request we ran. Configure gateway authentication according to the current Vercel documentation and keep credentials server-side. Confirm that your SDK version supports the selected gateway route and request features.

Start with a text-only smoke test, then add image input, structured output and tools independently. Inspect the response shape and usage records. A successful chat completion does not establish that every optional parameter works through every gateway. Our gateway architecture guide and OpenRouter guide explain the surrounding routing choices.

Pricing: the route is part of the number

Vercel announces 50% off through December 31, 2026. That is a dated offer on the gateway route, not a permanent promise about all Tencent endpoints. Check the actual price shown for your account and planned usage period before estimating a long-running application.

The OpenRouter listing, checked November 7, lists $0.417 input and $1.2505 output per million tokens, with $0.021 cached input where supported. Its current top-provider record allows up to 65,536 completion tokens. These are router figures, not a verified direct Tencent tariff.

At those OpenRouter rates, 100,000 uncached input tokens plus 10,000 output tokens costs $0.054205, before other applicable charges. That arithmetic does not represent a complete coding task. Repeated tool loops, retries and larger history can change the total substantially. The API pricing comparison covers the broader cost framework.

Do not apply Vercel’s 50% promotion to the OpenRouter figure. They are different routes. Keep provider, currency, verification date, promotion period and cache category alongside every stored price.

Vision and computer use

Tencent and Vercel describe visual understanding and computer-use training for graphical and command-line environments. A useful experiment combines an image with a constrained task, such as checking a chart against known source data or proposing a UI change from a screenshot.

Your application still provides the executor. Decide which tools the model may call, validate arguments and require approval for consequential actions. A screenshot can contain malicious instructions just as a web page can. Keep observations separate from trusted policy.

Do not equate a vision feature with video generation, browser access or reliable visual grounding on every interface. Measure the tasks you need, including ambiguous images and unavailable tools. The agent security guide explains why improved reasoning does not authorize broader access.

What a million-token context changes

The large window can accommodate more repository or document material. It does not guarantee that every relevant detail will be retrieved correctly, nor that full-window prompts are the cheapest approach. Remove generated files and duplicate evidence before comparing it with a smaller-context baseline.

Build an evaluation around answers with traceable evidence. Include a stale document, a conflicting requirement and a question whose answer is absent. Check whether the model identifies the correct source or invents a resolution. Our regression testing guide gives a repeatable structure for keeping that evidence useful across updates.

Self-hosting and alternatives

Until the final checkpoint and license are verified, use a hosted route for a capability trial rather than publishing a guessed GPU recipe. Preview hardware requirements do not establish the final model’s footprint. Recheck the official release channels for the artifact, license and supported runtimes together.

For alternatives, compare the released Mistral Large 4 and Kimi K3.1 Pro against the same workload. Their open-weight status does not make their licenses, memory requirements or behavior interchangeable.

My take

Hy4 is ready for a bounded hosted evaluation, especially when image input and long context belong in the same workflow. Keep the gateway discount separate from the steady-state budget. I would wait for verified final artifacts before planning a self-hosted purchase, while evaluating useful API work now.

FAQ

Is Hy4 the same as Hy4 Preview?

The current hosted release and the earlier preview are distinct release identities. Do not assume identical checkpoints or licenses.

Can I download the final weights now?

Tencent has announced them, but this check did not independently confirm a final download and license. The preview repository is not sufficient evidence.

Is the Vercel promotion permanent?

No. Its announcement specifies an end date of December 31, 2026.

Does computer use include an executor?

No. Your environment supplies tools, execution and authorization boundaries.