Tencent releases Hy4: the model reads 1M tokens, and API pricing starts at 83 cents

Tencent has released the final version of its large language model Hy4. The key numbers: it accepts up to 1M tokens per request (roughly an entire codebase or a stack of contracts), replies with up to 64K tokens, and costs 83 cents per 1M tokens of API input. The final version adds two capabilities: the model understands images, and it can operate interfaces on its own, both the command line and graphical windows across different operating systems. If you build agents that read documents or have to click through a screen where no convenient API exists, Hy4 is worth testing on your own tasks.

What the final Hy4 can do

Multimodality means the model accepts more than text. Hy4 recognizes images, parses documents and charts, and solves math problems presented as pictures. For example, you can send a screenshot of a report with a diagram, or a photo of a textbook page, and ask the model to explain what it shows.

Autonomous interface control means working in programs the way a human operator would. The model looks at the screen, finds a button or input field, types a command in a terminal, or clicks the right element. Agents that have to work with ordinary software need these skills.

The 1M-token input window lets you load a large repository or many documents into a single request, and the 64K-token output limit is enough for a long text or a large chunk of code.

Separately, Tencent says it removed offensive cyberattack tools from the training data. The company is trying to reduce the risk that the model will help write malicious code. How well this works will be shown by independent testing: cleaning the data alone does not guarantee that the model will refuse dangerous requests.

What it costs and how it's built

Hy4 has 770B parameters, but only 49B work on each token. Mixture-of-experts models are usually built this way: the network stores a lot of knowledge, but only part of it is computed per token. That makes such a model cheaper to run than a dense network of the same size.

Pricing: 83 cents per 1M input tokens and 2.5 dollars per 1M output tokens. The model is available through Tencent's services and on OpenRouter. With long context, input tokens drive the cost: a single request with a full million tokens comes to about 83 cents, and with frequent calls that sum adds up quickly.

What is still unknown and what to do

The model weights are still closed. Tencent promises to publish them in the coming days, followed by a lighter version, Hy4 Flash. Until that happens, you cannot deploy Hy4 yourself, and all your data has to go to a third-party provider. For companies with strict confidentiality requirements, that is a significant limitation.

It's best to verify the claimed capabilities (image handling, interface control, and long context) yourself. This applies especially to long context: formally the model holds a million tokens, but many models become less accurate at finding a specific detail in a volume that large.

The practical plan is simple. Put together a small set of real tasks: parsing a document with charts, an action in an interface, a question about a large file. Run them through Hy4 on OpenRouter and compare quality and cost with what you use now. If the weights are indeed published, self-hosting becomes an option, and Hy4 Flash may suit high-volume, low-cost requests.