Tencent Hy4 Opens 770B Weights: License Catch and Hardware Needs
Tencent has published the weights of Hy4, its 770B-parameter flagship model with 49B active parameters per token, image and video understanding, and a 1M-token context. It is not Apache 2.0: the Tencent Hy License adds limits for large hosting businesses. The smallest official GGUF build is about 215 GiB, so this is a multi-GPU server model, not a desktop one.
On this page
What Tencent put on Hugging Face this week
Tencent's Hunyuan team has published the full version of Hy4, the model it first showed as Hy4 preview in August. The Hugging Face organisation now carries three repositories: Hy4 (the chat-tuned model), Hy4-FP8 (an 8-bit floating point build of the same model) and Hy4-base (the pretrained model before chat tuning). A secondary report dates the open release to 20 November and describes it as completing the Hunyuan language model series, after a smaller Hy4 flash variant. The model card itself gives no release date.
The architecture is a sparse mixture of experts. The language backbone has 770B parameters across 78 layers, and each of the 77 expert layers holds 256 routed experts plus one shared expert, with 8 routed experts active for any token. That works out to about 49B active parameters per token. Compared with the preview, the card says Hy4 scales up both pre-training and post-training and adds a 1.23B vision encoder, so it accepts images and video as well as text. The context window is 1,048,576 tokens.
The Tencent Hy License is not Apache 2.0
Some early coverage, including the report cited above, says Hy4 ships under Apache 2.0. The license file in the repository says otherwise. It is the Tencent Hy License, Version 1.0, which states that it is derived from Apache 2.0 with modifications and "is not the Apache License".
The modification is a new Section 10 aimed at model-as-a-service businesses, meaning companies that give third parties remote access to inference or fine-tuning, for example through an API. Such a business needs a separate written license from Tencent if its worldwide gross revenue exceeds US$50 million in any 12-month period, or before using the model in a product with more than 100 million monthly active users. The rest of the grant keeps the familiar Apache terms: a perpetual, royalty-free license to use, modify and redistribute. Purely internal use is exempt from the hosting clause, and so are end-user products that confine the model to specific features without general-purpose access.
For individuals, researchers and most companies running the model on their own hardware, little changes in practice. But the license counts fine-tuned and quantized versions as derivative works, so the clause follows community builds too. Anyone who planned around the Apache 2.0 preview should read the new text before shipping.
How much hardware Hy4 actually needs
This is where the release stops being a local model in the everyday sense. Tencent's own compression team, AngelSlim, has published three GGUF builds: a standard 4-bit file at about 438 GiB, a mixed 2-bit and 4-bit file at about 310 GiB, and an experimental mixed-precision file at about 215 GiB. The README says keeping the whole model in GPU memory takes roughly those amounts of VRAM, plus the cache and about 3 GiB for the vision encoder, and suggests one node with eight 96 GiB accelerators.
The GGUF files are described as being for end-to-end testing and need a patched llama.cpp build, since the smallest one uses a quantization type that is not in mainline yet. For serving, the model card points to prebuilt vLLM and SGLang images, with examples that split the FP8 model across eight GPUs.
So the honest reading is this: Hy4 is self-hostable, which matters for organisations that want a frontier-scale model inside their own network, but it is far beyond a workstation. For a laptop or a single GPU, smaller open models remain the practical choice.
Benchmarks Tencent reports and their limits
All published numbers come from Tencent. On few-shot tests of the base model, the card lists 73.9 on MMLU-Pro against 73.4 for DeepSeek V4 Pro's base model, and 57.2 on SimpleQA against 52.1. It trails on the MBPP-plus coding test, 74.0 against 79.1. The card warns that base models sometimes break the expected answer format, which can distort scores.
For the chat model, Tencent reports 3.04 out of 4 on an internal finance evaluation, ahead of DeepSeek V4.1 Flash at 2.97, and says its measured hallucination rate fell from 15.76% in the preview to 5.45%. These are internal tests without outside replication. Tencent also writes that a "notable gap remains" between Hy4 and the best proprietary models on research-level maths, long-horizon coding and advanced visual question answering.