For sixteen days, the model was a promise. On November 4, Tencent unveiled the Tencent Hy4 model, the biggest AI system it had ever built, and told developers the weights would follow “in the coming days”. Then the days passed, the files never appeared, and the open-source label started to look like a slogan.
On Friday, the wait ended. Tencent has released the full weights of its flagship Tencent Hy4 model, according to TechNode, letting anyone download and run one of the largest open-weight AI models ever published. The files landed on Hugging Face in two formats, with no sign-up wall and no approval queue standing between developers and the download button.
It is a release that changes Tencent’s place in the AI race. Until now, the company’s flagship model was something you could only rent through an API. From today, you can own a copy.
What Tencent released on Friday
The release comes in two official checkpoints on Hugging Face: a full-precision BF16 version and a smaller FP8 version. Both carry the same model, a multimodal system that understands text, images and video, and both ship with a system card and safety evaluations.
The BF16 checkpoint weighs roughly 1.54 terabytes. The FP8 version is about half that, around 770 gigabytes. A community-built GGUF version has already appeared for researchers who want a quantized build for experimentation.
Tencent also published deployment instructions for vLLM and SGLang, fine-tuning resources and prebuilt Docker images tagged hy4. The company says the FP8 checkpoint can run across eight GPUs with tensor parallelism set to eight, and the server exposes an OpenAI-compatible API.
There is one surprise in the fine print. Early reports described the release as Apache 2.0, the licence that covered the August preview. The Hugging Face listing, however, shows a custom licence, the Tencent Hy License. Developers planning commercial use should read the repository’s licence text before deploying, not the headlines.
Tencent Hy4 by the numbers
Hy4 is a mixture-of-experts model, which means only a fraction of its brain switches on for any given word. The full model holds 770 billion parameters, but just 49 billion activate per token. That split is what makes a model this large affordable to run at all.
The backbone is a 78-layer transformer. The first layer is a conventional dense feed-forward network, while the remaining 77 layers are mixture-of-experts layers. Each of those layers holds 256 routed experts plus one shared expert, and for every token the router picks the top eight routed experts alongside the shared one.
The context window stretches to 1,048,576 tokens, roughly a million. That is enough room to feed the model a small library at once: hundreds of documents, a long codebase, or hours of video transcript, and still ask it to reason across all of it.
Vision comes from a 1.23-billion-parameter vision transformer that converts images and video frames at dynamic resolutions into tokens the same stack can read. The result is a single model that can look at a screenshot, read a document and operate a computer interface.
A million-token memory without the usual bill
A million-token context sounds impressive until you pay for it. In a normal transformer, the cost of attention grows with every token you add, so a million-token window can turn into a very expensive conversation.
Hy4 tackles that with Gated DeepSeek Sparse Attention and a component called IndexCache. An indexer picks the 2,048 positions most relevant to each token, and the cache reuses those selections across layers instead of scanning the whole context again and again.
The model also ships with DFly2, a speculative decoder built for the sparse-attention system. Each draft layer shares its corresponding Hy4 layer’s key-value cache and selected positions, so draft attention looks at roughly 2,048 keys per token whatever the context length. Tencent says the drafter needs no extra key-value cache memory.
The practical payoff is speed. The model can also run in a direct-response mode with no extended thinking, or switch to a high reasoning effort for harder problems, using a parameter called reasoning_effort. Native tool calls are available when the hy_v4 reasoning and tool-call parsers are switched on.
What the benchmarks say, and what they do not
Tencent’s own numbers put Hy4 ahead of its Chinese rivals on most knowledge and reasoning tests. On MMLU it scores 91.6, on MMLU-Pro 73.9, and on the graduate-level SuperGPQA benchmark 62.7. Factual recall looks strong too: SimpleQA comes in at 57.2, up from 26.4 for the Hy3 preview.
Multimodal results are solid as well. MMMU-Pro, a test of multimodal reasoning, lands at 59.6. Visual understanding scores 86.7 on CVBench, and visual counting reaches 90.6 on CountBenchQA. The company says Hy4 leads its comparison with DeepSeek-V4-Pro-Base and DeepSeek-V4-flash-Base on most of these tests.
But every one of these figures comes from Tencent itself, and none has been independently reproduced yet. That matters, because the same tables also reveal a weakness. On MBPP-plus, a code-generation test, Hy4 scores 74.0, trailing DeepSeek-V4-Pro-Base at 79.1. Coding is the clearest gap in the public results.
Tencent also reports weaker performance against closed models on research-level mathematics, long-horizon coding, model self-improvement and advanced visual question answering. And the published results do not establish how well the model retrieves information or reasons across the full million-token window, which is the feature the release is built around.
Built for real work: finance, games and computer use
Where Tencent claims the clearest edge is in specialist work. The company evaluated Hy4 on 579 financial tasks with 213 finance professionals, and says it ranked first among the open-weight models tested, ahead of DeepSeek-V4.1-Flash, Kimi K3.1 Pro and GLM-5.3.
Hallucination is another area Tencent highlights. Internal evaluations cut the reported hallucination rate from 15.76 percent in the Hy4 preview to 5.45 percent in the finished model. A category the company calls “speculative errors” fell from 23.1 percent to 5.6 percent.
Game development got targeted training too. Working with developers from Tencent’s own games division, the team tuned the model for Unreal Engine, Unity and Godot workflows, reporting gains of 19.1 percent on GameDev and 9.8 percent on GameLogic over the preview model.
Then there is computer use, the capability that separates Hy4 from its preview. The model was trained on Blender, CAD software, command-line tools and graphical interfaces, learning to plan across long, multi-application workflows. It is the first model in Tencent’s Hunyuan line that can both see images and operate a computer, reading screens, clicking interface elements and running commands.
Why this release matters now
Hy4 arrives in the middle of a Chinese open-weight surge. DeepSeek, Moonshot’s Kimi and Zhipu’s GLM have all shipped large open models this year, Meituan’s LongCat-2.5 arrived under the permissive MIT licence, and Alibaba’s Qwen4-Max remains preview-only, with no weights in sight. Tencent is now firmly in that group, and with the full flagship rather than a cut-down version.
The timing is deliberate. Tencent already opened the smaller Hy4 Flash, a 224-billion-parameter sibling, earlier this month. Releasing the flagship weights now gives developers the complete family: a budget model for everyday work and a flagship for the hardest jobs.
For developers who do not want to host 770 gigabytes of weights, the hosted route stays open. Hy4 is available through Tencent Cloud TokenHub, inside CodeBuddy and WorkBuddy, in the Yuanbao app, and through third-party APIs including OpenRouter, where it lists at $0.834 per million input tokens and $2.501 per million output tokens.
What happens next
The next test is not Tencent’s. It belongs to the community. Independent benchmarks will show whether the million-token window holds up under real workloads, whether the coding gap is real, and how the model behaves once it is fine-tuned for jobs Tencent never imagined.
Watch for the first community fine-tunes and the first GGUF builds running on smaller hardware. Watch, too, for Artificial Analysis and other independent evaluators to publish their own numbers. Until then, Hy4’s scores are a claim, not a verdict.
What is no longer in doubt is the download itself. Sixteen days after the announcement, the weights are real, the licence is Tencent’s own, and the model that was only a hosted API is now a file anyone can copy. For the open-weight movement, that is the part that counts.



