腾讯混元

Hunyuan is the brand Tencent, the Chinese technology company that owns WeChat, uses for its in-house foundation models. A foundation model is a large neural network trained on broad data that can then be adapted to many tasks. The Hunyuan family is not one model but several lines: general large language models for text and code, image generation models, the HunyuanVideo line for video and the Hunyuan3D line for 3D assets. Recent model cards and announcements use the shorter name Hy, as in Hy3 and Hy4, and describe the developer as the Tencent Hy team. The models serve two audiences: Tencent's own products, such as the Yuanbao chatbot and the CodeBuddy and WorkBuddy assistants, and outside developers through Tencent Cloud and open-weight downloads.

The first public milestone came on 7 September 2023, when Tencent demonstrated Hunyuan in Shenzhen and opened it to enterprise customers. Reuters reported that Tencent described the model as having more than 100 billion parameters, the adjustable values a network learns in training, and as trained on more than 2 trillion tokens, the word fragments a language model reads. Tencent claimed Hunyuan hallucinated, meaning produced confident false statements, 30 percent less than Meta's Llama 2, and said it already underpinned more than 50 internal products; Reuters noted these claims were not independently verified. The launch came in a crowded field: Tencent vice president Jiang Jie said China already had more than 130 large language models by July 2023.

In November 2024 Tencent made its most significant early open release, Hunyuan-Large. Its arXiv paper describes a mixture-of-experts (MoE) model with 389 billion total parameters but only 52 billion activated for each token. In an MoE model, each layer contains many smaller expert sub-networks and a router picks a few of them for each token, so the model can store a lot of knowledge while spending compute like a much smaller model. The paper reports a context window of up to 256,000 tokens, heavy use of synthetic training data and compression of the key-value cache, the memory a transformer keeps about earlier tokens, to make long inputs cheaper. Code and checkpoints were released publicly.

Tencent also built a strong presence in open visual generation. HunyuanVideo, posted in December 2024, is a video generation model with over 13 billion parameters, which its authors described as the largest open-source model of its kind at the time; in professional human evaluations they report it outperformed Runway Gen-3, Luma 1.6 and three leading Chinese video models. Hunyuan3D 2.0, posted in January 2025, splits 3D generation into two models: Hunyuan3D-DiT, a flow-based diffusion transformer that generates a bare shape matching a condition image, and Hunyuan3D-Paint, which generates high-resolution texture maps for that mesh or for one made by hand. A diffusion transformer is an image or shape generator that starts from noise and removes it step by step under the control of a transformer network. The authors report it beats earlier open and closed systems on geometry detail and texture quality, though these are self-reported results.

The language-model line was then rebuilt. Hy3 preview, open-sourced on 23 April 2026, is described on its model card as the first model trained on the team's rebuilt infrastructure. Its model card lists 295 billion total parameters with 21 billion active, 192 experts of which 8 are active per token, a 256,000-token context window and a choice of reasoning effort from none to high. It reports 87.2 on the GPQA Diamond science benchmark and 74.4 percent of tasks resolved on SWE-bench Verified, a test of fixing real software issues. The final Hy3 launched on 6 July 2026, with more post-training compute and better data according to Caixin, which also reported Tencent's claim of a 90 percent completion rate on agent tasks in internal applications.

Hy4 preview followed on 28 August 2026. Tencent's announcement lists 770 billion total parameters, 49 billion active and a context window above one million tokens, with API prices of about 0.83 US dollars per million input tokens and 2.50 dollars per million output tokens. The Hugging Face model card describes 78 layers, 256 routed experts plus one shared expert with the top 8 routed experts used per token, and an Apache 2.0 license, a permissive open-source license that allows commercial use. It reports scores such as 92.3 on GPQA Diamond and 65.7 on SWE-bench Pro. In an internal blind test of 203 engineering tasks judged by 163 experts, Tencent reported an average rating of 2.99 out of 4, against 2.92 for GLM-5.3 and 2.94 for Kimi K3.

Compared with other Chinese model families, including GLM and Kimi, which Tencent itself uses as reference points, Hunyuan stands out less for a single architectural idea than for breadth: Tencent maintains competitive language, video and 3D lines at the same time and ties them to products with very large user bases. Its recent language models follow the same broad recipe as many peers, large sparse MoE transformers with long context and adjustable reasoning, and the Hy4 card names DeepSeek Sparse Attention as part of its design. Licensing has varied by release, which matters for builders: Hy3 preview uses the Tencent Hy Community License, while Hy4 preview uses Apache 2.0.

The main limitations are about verification and access. Most headline numbers, from the 2023 hallucination claim to the Hy4 blind evaluation, come from Tencent itself, and the comparison models are chosen by Tencent. Benchmarks such as GPQA and SWE-bench measure narrow skills and can be affected by training data overlap. Running the largest open models is expensive: Tencent's own deployment examples for Hy3 and Hy4 use eight large-memory GPUs, and 295 billion parameters stored in 16-bit precision alone take roughly 590 GB. For most users, practical access is through Tencent's apps or paid APIs rather than self-hosting.