Tencent Hy4: what it is, pricing, how to use it

Tencent Hy4 is the full release of Tencent's flagship AI model, launched November 4 with image input and computer use added to the August preview. List price is $0.834 per million input tokens and $2.501 per million output tokens, and you can try it now through OpenRouter, Tencent Cloud or Tencent's own apps.

The launch post from the @TencentAI_News account has one good line: "The price stayed in preview. The benchmarks didn't." The same post says open weights are coming soon, with Hy4 Flash right behind.

Hold on to that word "coming." Hy4 is being described as an open model, and today it isn't one. @arena put it plainly when it ranked Hy4 #14 overall in Agent Arena: "Currently proprietary." Arena adds that once the weights are out, Hy4 would tie Kimi K3.1 Pro for first among open models.

What Hy4 can do

Two things are new compared with the preview. With vision it can read image details and complex charts and do visual math, according to ChainCatcher's launch report. Computer use lets it carry out tasks across both the command line and graphical interfaces.

The OpenRouter route has a context window of 1,048,576 tokens and a 65,536 token output limit, per Digital Applied's rundown of what is confirmed. The same rundown notes Tencent hasn't stated a parameter count or license for the final model. The 770 billion total and 49 billion active figures you'll see quoted are documented for the preview on GitHub, which is the only Hy4 you can download today, under Apache 2.0.

Tencent Hy4 pricing

The OpenRouter rate is a 50% promotion with no stated end date. ChainCatcher reports a matching half-price offer on Tencent Cloud TokenHub through December 31, 2026. For comparison, Kimi K3.1 Pro lists at $2 and $10.

How to use Tencent Hy4

1. Fastest for developers: create an OpenRouter key and call the model ID tencent/hy4.
2. Tencent Cloud: use it directly if you already have an account there.
3. Tencent's apps: Hy4 is also in CodeBuddy, WorkBuddy, Yuanbao and ima. ChainCatcher reports free usage in WorkBuddy and CodeBuddy from November 4 to November 15. None of the coverage we read says which of these apps work outside China, so check before planning around them.
4. Self-hosting: only the preview is available, and its example vLLM and SGLang commands use eight-way tensor parallelism, so this is a job for a GPU server.

Here is a first test through OpenRouter's OpenAI-compatible API. We haven't run it ourselves. The listing says the route accepts images, but none of the coverage we read reported testing that.

from openai import OpenAI

client = OpenAI(
    base_url="https://openrouter.ai/api/v1",
    api_key="YOUR_OPENROUTER_KEY",
)

response = client.chat.completions.create(
    model="tencent/hy4",
    messages=[{
        "role": "user",
        "content": [
            {"type": "text", "text": "Read this chart. List the three biggest changes and any number you can't read clearly."},
            {"type": "image_url", "image_url": {"url": "https://example.com/your-chart.png"}},
        ],
    }],
)
print(response.choices[0].message.content)

That last clause in the prompt is worth keeping. It shows you whether the model admits what it can't see or fills the gap with a guess.

Whether computer use works through the API isn't clear from anything we could read. If that's the feature you want, Cursor Computer Use is a useful point of comparison for what to test.

How far to trust the numbers

Most of the launch figures come from Tencent. In a Tencent-run finance test with 213 professionals and 579 tasks, Hy4 scored 3.04, ahead of DeepSeek V4.1 Flash at 2.97 and Kimi K3.1 Pro at 2.68. TechNode, citing Tencent's materials, puts Hy4 13th on Agent Arena and at about $0.78 per task on an Artificial Analysis benchmark. Arena's own post said 14th. The gap is small, and I couldn't tell whether the two figures were taken on the same day.

My read: the demand looks real, and the "open" label is ahead of the facts. TechNode reports that Tencent expanded Hy4's inference capacity tenfold on November 5 after WorkBuddy users hit their limits on day one. The weights have no date yet, though. Don't plan on self-hosting the final model until they ship.

What this means for a business

A model at this price that can also read images makes it easier to justify automating work like pulling numbers from invoices and screenshots. The per-token price shouldn't settle the choice on its own, though. On the OpenRouter page Tencent Cloud is the only named provider, so ask where your data goes before you send it customer records. Then run one real task on two models and compare the output and the bill.

If you'd like help working out which of your processes are ready for that kind of test, our free process audit starts there.