Tencent Unveils Low-Cost Hy4 AI Model With 770bn Parameters
The Tencent Hy4 AI model is an open-weight system built to compete with leading models from OpenAI, Anthropic, DeepSeek, Zhipu AI and Moonshot AI. Tencent released Hy4 Preview on August 28, 2026, and the model has drawn fresh attention for its low API prices and its strong scores on selected coding tests.
Open weights set Hy4 apart from closed rivals such as GPT-5.6 and Claude Opus 5. Developers can download the weights and run the model on their own systems, instead of relying only on a company’s servers. That choice gives teams more control over cost, privacy and setup, and it suits firms that want to build their own tools on top of a large model.
Hy4 Preview specifications
Hy4 Preview has 770 billion total parameters, but it activates only 49 billion parameters for each token. It uses a Mixture-of-Experts design, which means it calls on a small set of specialist parts for each step, rather than the whole model at once. The design keeps running costs lower than the total size suggests.
The model supports a one-million-token context window. In plain terms, it can take in a very large set of documents, code files or records in one go. That helps with long coding tasks, document review and research work, where the model needs to keep many details in view at once.
Tencent built Hy4 with 78 layers and a multi-token prediction layer that helps it plan ahead as it writes.
| Detail | Hy4 Preview |
|---|---|
| Total parameters | 770 billion |
| Active parameters per token | 49 billion |
| Context window | 1 million tokens |
| Layers | 78 |
| Licence | Apache 2.0 |
| Release date | August 28, 2026 |
Low API pricing
Price is the second reason Hy4 stands out. Tencent lists cached input at $0.042 per million tokens, standard input at $0.834 per million tokens, and output at $2.501 per million tokens. Cached input costs far less, since the system can reuse work it has already done.
For a small team, that pricing changes the maths. A firm that runs many coding or document tasks can test ideas without a large upfront bill. For a large firm, the saving grows with volume. However, price alone does not decide quality, and teams should still test the model on their own tasks before they commit to it.
How Hy4 scores on coding tests
Tencent’s own benchmarks put Hy4 ahead of several rivals on selected coding tests, but not on every test. On SWE-bench Multilingual, Hy4 scored 82.9, ahead of DeepSeek V4 Pro at 77.3, GLM 5.3 at 81.3, Kimi K3 at 80.8 and GPT-5.6 Sol at 74.1. Claude Opus 5 scored higher on that test, at 89.5 or 85.8 under different setups.
What the Tencent Hy4 AI model is built to do
Tencent designed Hy4 for four kinds of work: software engineering, office productivity, game development and scientific research. In coding, it handles code reading, planning, debugging, testing and front-end work. In office work, it reads documents, spreadsheets and presentations, and it can check many files at once.
Tencent also lists two limits. The model can spend too long reasoning through hard tasks, and it can check its own work more often than needed. Both habits raise response times and costs. Tencent released the preview early to gather feedback, and it plans further training to fix those issues.
Why it matters for developers in Pakistan
For developers in Pakistan, the open weights and low prices matter more than any single benchmark. A team that cannot afford high API bills can download Hy4 and run it on rented hardware, or use the low-cost API for early tests. That lowers the barrier for startups, university labs and small software houses that want to build AI tools for local clients.
The one-million-token window also suits local needs. A firm that handles long legal files, medical records or government forms could feed large sets of documents to the model in fewer steps. However, teams must still check the output, since benchmark scores do not guarantee correct answers on real client work. Data privacy rules, client consent and human review remain essential, even with an open model.
The wider AI race also shapes the context. Feature Pakistan reported on warnings that AI systems may soon build parts of themselves in Anthropic Warns AI May Soon Build Itself, and on Apple’s plans for AI glasses in Apple AI Smart Glasses Tipped for Late 2027. Hy4 fits the same trend: large models keep getting cheaper to use, even as their makers warn that human checks still matter.
ProPakistani reported the Hy4 details on November 4, 2026, drawing on Tencent’s published benchmarks.
Frequently asked questions
What is the Tencent Hy4 AI model?
It is Tencent’s open-weight AI model, released as Hy4 Preview on August 28, 2026, with 770 billion total parameters.
How many parameters does Hy4 activate per token?
It activates 49 billion parameters per token, using a Mixture-of-Experts design.
What does open-weight mean?
It means developers can download the model weights and run the model on their own systems, instead of using only a company’s servers.
How much does Hy4 API access cost?
Tencent lists cached input at $0.042, standard input at $0.834 and output at $2.501 per million tokens.
Does Hy4 beat every rival model?
No. It leads on selected coding tests, but rivals such as Claude Opus 5 score higher on some benchmarks.
Where can developers access Hy4?
They can use Tencent Cloud, OpenRouter and Tencent’s AI applications, or download the weights from open-source repositories.
