On August 28, 2026, Chinese technology giant Tencent officially introduced its new flagship model, Hy4 Preview, releasing it under the permissive Apache 2.0 open-source license. The foundation model relies on a Mixture-of-Experts (MoE) architecture comprising a total of 770 billion parameters, with 49 billion parameters actively routed per token during inference. A defining technical characteristic of the model is its context window spanning 1 million tokens, engineered to handle extensive codebases and complex multi-document inputs in a single pass.
From an architectural perspective, Hy4 Preview implements Gated DSA-Attention alongside a dedicated 10-billion-parameter Multi-Token Prediction (MTP) layer. This MTP layer is designed for speculative decoding, substantially accelerating inference speeds during token generation. To facilitate immediate deployment for developers, Tencent made the model weights directly available on Hugging Face while simultaneously releasing production-ready Docker containers for inference engines such as vLLM and SGLang, complemented by specialized quantization toolkits.
According to Tencent, Hy4 Preview was trained collaboratively alongside internal engineering teams specializing in software development, financial analysis, and gaming. Notably, the model was deployed during development to assist in autonomous infrastructure refinement. Hy4 Preview independently optimized segments of its own inference pipeline and operator fusion routines, which Tencent reports delivered an operational throughput improvement of 31.8 percent.
In internal blind evaluations focused on programming benchmarks, Hy4 Preview narrowly outperformed competing models such as GLM-5.3 and Kimi K3. Tencent is positioning the open-weights system for long-horizon workflows, autonomous agentic coding, and scientific research. Targeted scientific applications include computationally intensive tasks within molecular dynamics and solid-state physics, where reasoning over vast context windows is required.
The release of a 770-billion-parameter open-weights model under Apache 2.0 highlights the intensifying rivalry between open-source initiatives and proprietary AI providers. The pressure exerted by frontier open-weights releases has already forced commercial vendors to adjust, as reflected in OpenAI reducing API pricing for its GPT-5.6 Sol model by over 20 percent. By pairing open weights with turn-key deployment images for vLLM and SGLang, Tencent is streamlining the adoption of frontier AI in enterprise production environments.