Tencent on November 20, 2026 released the open weights for Hy4, the flagship model from its Hy (Hunyuan) team. According to the official Hugging Face model card, the release includes the instruction-tuned Hy4, an FP8-quantized Hy4-FP8, and the pre-trained Hy4 base model, published on Hugging Face, ModelScope and GitCode.

Hy4 is a mixture-of-experts model whose language backbone has 770 billion total parameters, of which 49 billion are activated per token. Tencent lists 78 layers, with 256 routed experts plus one shared expert per MoE layer and eight routed experts active for each token. A 1.23 billion-parameter vision encoder lets the model handle images and video alongside text, and the context window is 1,048,576 tokens.

The model card says the architecture borrows from DeepSeek and GLM designs, using Gated DeepSeek Sparse Attention with an IndexCache for reusing sparse indices across layers. Tencent also built in a 3.3 billion-parameter drafter, called DFly2, for speculative decoding, which reuses Hy4's key-value cache to speed up generation on long contexts.

Compared with the earlier Hy4 preview, Tencent says it scaled up pre-training and post-training, added vision, and improved instruction following, long-context understanding, long-horizon tasks and deep research. The company highlights computer-use workflows across command-line and graphical interfaces, front-end and game development, and finance and office tasks, and reports gains of 19.1% on its GameDev evaluation and 9.8% on GameLogic over the preview. Those figures are Tencent's own.

Tencent is also candid about limits. The model card states that "a notable gap remains" between Hy4 and top proprietary models, particularly on internal benchmarks for research-level mathematics, long-horizon coding, model self-improvement and advanced visual question answering.

Licensing is the detail developers should read closely. Hy4 ships under the Tencent Hy License, Version 1.0, which the license file describes as derived from Apache 2.0 but explicitly "not the Apache License." Its added Section 10 requires companies that offer the model as a general-purpose service, such as through an API, to obtain a separate written license from Tencent if their worldwide revenue exceeds $50 million over 12 months or if the product has more than 100 million monthly active users. Purely internal use is exempt. Some early coverage described the release as Apache 2.0, which the license text does not support.

For deployment, Tencent provides instructions for vLLM and SGLang with an OpenAI-compatible API, along with fine-tuning resources and a system card covering safety evaluation. The model card does not list hosted API pricing.