Tencent launches Hy4 Preview with open weights and low API costs

A new open-weight contender

Tencent has officially released Hy4 Preview, a large-scale AI model designed to compete with major systems from companies like OpenAI and Anthropic. Unlike many proprietary models, Hy4 Preview is an open-weight model, meaning developers can download the weights and deploy the system independently. This release provides a high-performance option for those who want more control over their AI infrastructure.

The model is built with 770 billion total parameters, though it uses a Mixture-of-Experts (MoE) architecture to manage resources efficiently. This specific design allows the model to activate only 49 billion parameters for each token it processes. By focusing on specific experts for different tasks, the system maintains high performance while managing the computational demands of its large size.

Tencent positioned this release as its most significant generation-over-generation improvement to date. The company aims to provide a model that excels in software engineering, office productivity, and scientific research. It is currently available through Tencent Cloud, OpenRouter, and various open-source repositories under the Apache 2.0 license.

Technical architecture and specs

The internal structure of Hy4 Preview includes 78 layers, with 77 of those being MoE layers containing 256 routed experts. Each layer also features a shared expert to ensure consistent processing across different types of inputs. To improve how the model handles information, Tencent integrated Gated DeepSeek Sparse Attention and identity Hyper-Connections into the architecture.

One of the standout technical specifications is the model's one-million-token context window. This allows the AI to process and remember vast amounts of information in a single session, which is particularly useful for analyzing long documents or large codebases. Tencent achieved these capabilities by increasing the model size and the volume of training data used during both pre-training and post-training phases.

The company also focused on making the model accessible to different hardware environments. In addition to the standard version, Tencent provided an FP8 version of the model. This version is compatible with popular deployment tools like vLLM and SGLang, making it easier for developers to integrate into existing software pipelines.

Benchmark performance and testing

Tencent reported that Hy4 Preview shows significant gains over its predecessor, Hy3. In the DeepSWE benchmark, the score jumped from 28.0 to 64.3, while the SWE Atlas Codebase Q&A score improved from 30.8 to 64.0. These results suggest that the new model is much more capable of handling real-world engineering tasks and complex multi-step workflows than previous versions.

When compared to other frontier models, Hy4 Preview showed mixed but competitive results. It outperformed several rivals on the SWE Atlas Codebase Q&A test, but it still trailed behind Claude Opus 5 on the SWE-bench Multilingual test. On the Terminal-Bench 2.1, it scored 85.4, though several other models reported higher results in that specific category.

To further validate the model, Tencent conducted a blind evaluation involving 163 internal experts who reviewed 203 software engineering tasks. Hy4 Preview received an average score of 2.99 out of 4, slightly edging out competitors like GLM 5.3 and Kimi K3. While these are internal results, they indicate that the model is performing at a high level in practical coding scenarios.

Practical applications for users

Tencent developed Hy4 Preview by incorporating direct input from various professional fields. Software engineers helped shape its ability to handle long development tasks and front-end design. For example, the model was used to demonstrate a 3D viewer that followed specific requirements for layout, typography, and interactive controls.

In the realm of office productivity, the model was tested on a workflow involving 24 reimbursement claims. It had to analyze multiple policy versions, staff records, and invoices to prepare final decisions. For game development, the model was used to create a prototype for a cartoon adventure game, including elements like inventory management, quest information, and sound effects.

The model also shows promise in scientific research, particularly in fields like molecular dynamics and pure mathematics. In one experiment, Hy4 managed several Codex sessions simultaneously to coordinate experiments and adjust research directions. While it cannot independently develop new frontier AI systems, it demonstrated a high capacity for managing complex, multi-step research workflows.

Pricing and known limitations

To make the model accessible for large-scale business use, Tencent introduced a usage-based API pricing structure. Standard input costs $0.834 per million tokens, while output costs $2.501 per million tokens. However, the company offers a significantly lower rate of $0.042 per million tokens for cached input, which allows users to reuse previously processed information more cheaply.

Despite its high performance, Tencent acknowledged that Hy4 Preview still faces certain technical hurdles. The company noted that the model sometimes spends too much time reasoning through complicated tasks or repeatedly checking its own work. These behaviors can lead to increased response times and may impact the overall efficiency of the system during heavy use.

Tencent released this preview version early specifically to gather feedback and identify these types of issues. The company plans to continue refining the model through additional training and collaboration with internal teams. For now, it serves as a powerful, low-cost tool for developers who need a high-parameter model with open weights.