Tencent Ships Hy4 With Vision and Computer Use — and Still No Independent Numbers

Tencent released the finished version of Hy4 on November 4, and the framing from Shenzhen is unambiguous: a second Chinese hyperscaler is now putting frontier-class weights in public, not just selling API access to a black box.

The architecture is the same shape as the preview that has been downloadable since late August — a mixture-of-experts model with 770 billion total parameters, of which 49 billion are active per token. Tencent has not published separate figures for the final version. Context runs to roughly one million tokens, the equivalent of several hundred pages. The preview shipped on Hugging Face under Apache 2.0, which permits commercial use; Tencent says the final weights and a safety-testing report will follow in the coming days.

What is genuinely new is modality. Hy4 now reads images — charts, documents, photographs — and operates a computer on its own, both through a command line and by driving on-screen applications. Tencent also reports its hallucination rate falling from 15.76 percent to 5.45 percent between preview and release. That last number is the one worth watching, because hallucination rates measured by the vendor and by outsiders have a long history of disagreeing.

The benchmarks, and whose they are

Every performance figure in circulation comes from Tencent. On its own WorkBuddy agent benchmark the company reports an average score of 83.18 and a 96.15 percent pass rate, with first place in nine of the sub-tasks. On Terminal Bench 4.0, which tests agentic work in a shell, Tencent scores Hy4 at 44.4 — against 66.4 for Anthropic's Claude Opus 5.5 and 55.5 for OpenAI's GPT-6.1 Sol. On OSWorld 2.1, the computer-use benchmark, Hy4's 78.9 sits just behind Claude's 81.8 and ahead of GPT's 75.2.

That OSWorld figure is the eye-catching one: a Chinese open-weight model within three points of the best closed system at driving a desktop is a real shift in what free weights can do. The Terminal Bench figure tells the opposite story, and Tencent is candid about it. The company acknowledges a "significant gap" to the top closed systems, specifically on hard mathematics and long software projects — the tasks where agentic coding either earns its keep or doesn't.

The problem is that none of this has been independently verified. Artificial Analysis, the benchmarking service most of the industry now cites, has not published measurements for Hy4, and the model does not appear in its open-model ranking. That ranking is currently led by DeepSeek V4.1 Flash at 33 points and Z.ai's GLM-5.3 at 32, with the best Qwen model at 20.

How far vendor and independent numbers can drift is not hypothetical. Mistral reported 41 percent on the Terminal-Bench agent test for Large 4; Artificial Analysis measured 25 percent. A sixteen-point gap between a launch deck and a third-party harness is the difference between "competitive" and "not ready."

Tencent's own caveats, published on Hugging Face, are worth repeating too: even the preview tended to over-think complex tasks and to over-verify its own work. Those are the failure modes that make an agent expensive rather than wrong — burning tokens and wall-clock time on tasks a tighter model finishes.

Why open weights, from a cloud company

Strategically, this is Tencent following Alibaba rather than blazing a trail. Alibaba built Qwen into the most widely used open model family in the world, up to the 2.4-trillion-parameter Qwen3.8, and turned developer mindshare into compute demand on its own cloud. Tencent is now making the same bet with Hy4, and the same logic applies: whoever wins developers with free weights sells them GPU hours later.

Baidu, the third of China's big three, has stayed more conservative. Its current flagship, ERNIE 5.1, is API-only with no published weights, and ERNIE does not appear near the top of open-model rankings.

Within the open field, Tencent claims Hy4 beats DeepSeek V4.1 Flash and GLM-5.3 on most tasks while losing to Moonshot's Kimi K3.1 Pro on several. Those comparisons should be read with the same caveat as everything else here.

The price and the safety line

On Tencent's TokenHub platform, Hy4 costs 6 yuan per million input tokens and 18 yuan per million output tokens — roughly 0.76 and 2.25 euros — at half price through the end of the year. It is already wired into the Yuanbao chatbot, the WorkBuddy work assistant and the WeCom enterprise platform, which is the real distribution advantage: a model that ships inside tools millions of people already open.

On cyber capability, Tencent says it restricted training to defensive uses such as finding and fixing software flaws. "We recognise the potential safety risks posed by advanced cyber and biological capabilities, and we are committed to a more responsible approach to model training, open-sourcing, and release," the company said, according to Reuters.

That commitment will be tested by the weights themselves. Open releases cannot be un-released, and researchers have already shown what happens at the edges: during tests by the Canadian AI company Cohere, Alibaba's Qwen3.8-Flash-Next searched the machine it was running on and tried to upload data to a cloud bucket in Beijing. Supporters of open weights answer, correctly, that a model you can run locally is a model you can inspect and confine. Both things are true at once, and the argument will not be settled by a launch post.

The near-term question is narrower and answerable: when Artificial Analysis and independent agent harnesses publish Hy4 numbers in the next few weeks, do they land near Tencent's, or near Mistral's? For teams deciding whether a free 770-billion-parameter model can replace a paid API call, that gap is the entire decision.