Tencent's Hunyuan 4 flagship ships as an Apache 2.0 open model, with no territory clause

Tencent released Hy4-preview, the flagship of its Hunyuan 4 line, on 28 August 2026, with open weights on Hugging Face, ModelScope, GitCode and CNB. We are covering it now because the detail that matters most for a UK reader is one we have been writing about all month, and this release lands on the right side of it.

The news

Hy4-preview is a 770-billion-parameter text mixture-of-experts with about 49 billion parameters active per token: 256 routed experts and one shared expert across 78 layers, with the router selecting eight routed experts for each token. It carries a 1-million-token context window and a reasoning mode that defaults to deep chain-of-thought and can be turned off. Tencent published it in both BF16 and FP8 formats. It is text-only, with no multimodal input or output. Tencent positions it for coding, office productivity, game development, and scientific research; those capability claims are the company’s own and await independent testing.

The licence, read at source

The reason this is worth a piece rather than a catalogue line is the licence. Earlier this month we wrote about how several open models exclude UK users by licence, and Tencent’s own Hunyuan media models were a central example: our catalogue entry for HunyuanVideo carries a licence whose territory excludes the European Union, the United Kingdom and South Korea, and that exclusion reaches the model’s outputs. Because of that history, and because secondary write-ups have repeatedly called territory-restricted Hunyuan models “Apache 2.0” when they were not, we read Hy4-preview’s licence at the source rather than trusting the summaries.

It checks out. The LICENSE file in the Hy4-preview repository is the standard Apache License 2.0, word for word, with no “Territory” definition and no exclusion of the EU, UK or South Korea anywhere in the text. It is a worldwide, perpetual, commercial grant. In plain terms, a UK reader can download it, run it, fine-tune it and use its outputs, including on-premises and commercially, on the same footing as anyone else. As always, read the licence yourself before relying on it; we are describing it, not giving legal advice.

Why it matters

This is the clearest illustration yet of the point our earlier piece was making: within a single developer, the licence varies model by model, and the word “open” tells you nothing about the territory clause. Tencent ships a video model a UK user is not licensed to touch and, weeks later, a flagship language model a UK user can use freely. The only way to know which is which is to read each licence. For anyone who held off the Hunyuan family on licence grounds, this LLM is a different proposition from the media models, and a welcome one.

Set the licence aside and it is also a significant release in its own right. A 770-billion-parameter open mixture-of-experts under Apache 2.0 puts Tencent’s flagship into the same genuinely-open tier as DeepSeek V4.1 Flash, and a step more permissive than the bespoke licence on the GLM-5.3 flagship. The run of very large open models continues, and the licences attached to them are, on balance, getting more permissive rather than less.

Who it is for, and the catch

For local use, be clear about scale. At 770 billion parameters this is server-class: even at a 4-bit quantisation the weights alone need on the order of 385GB of memory, and at the released FP8 precision closer to 770GB, so this is a multi-GPU server rather than a workstation. Those are estimates, and the real figure depends on the quantisation and serving stack. Production GPU servers such as vLLM and SGLang are the realistic route to running it. The sparsity helps the economics once you are at that scale: only about 49 billion of the 770 billion parameters compute per token, so it serves more like a mid-sized model than its total suggests, which is the same pattern we keep seeing across this class of release. The reasoning mode is on by default and adds latency and token use, so turn it down for straightforward tasks.

We have added Hy4-preview to the model catalogue with its confirmed Apache 2.0 licence and hardware requirements, and we will fill in independent benchmark results as they appear.