Tencent Opens the Weights of a 770B Model — US Labs Close, Chinese Labs Open

Tencent has released Hy4 preview, a next-generation large language model, and opened its weights. The model carries 770 billion total parameters with 49 billion active, and a context window exceeding 1 million tokens. That mixture-of-experts design keeps the model large while activating only about 6% of it per forward pass — scale without the inference bill that usually comes with it.

Tencent aimed the model at four categories of real work: coding, office productivity, game development and scientific research. It reports stronger understanding, debugging and validation on long-context development tasks, support for workflows running from information processing through to documents, spreadsheets and presentations, and the ability to turn a single natural-language request into a playable game prototype that connects to game engines. The model is reachable by API through Tencent Cloud TokenHub and OpenRouter, and runs inside Tencent’s own WorkBuddy, CodeBuddy, Yuanbao and ima products.

What “Opening the Weights” Actually Means

An AI model splits into two parts. The architecture — how many layers, wired how — is published in papers and anyone can rebuild it. The weights are the numbers that fill that structure in, tuned incrementally through training. In Hy4’s case there are 770 billion of them. The architecture alone does nothing; the intelligence lives entirely in those numbers. Training is the process of spending hundreds of millions of dollars in compute to settle their values, and the weights file is the output.

What Tencent calls “open-sourcing” is, strictly, open-weight. The tencent/Hy4-preview repository on Hugging Face contains 131 weight files and fine-tuning scripts under an Apache 2.0 license permitting commercial use — but no pretraining data and no training recipe. You get the finished brain without being told how it was built. DeepSeek, Qwen and Llama all work the same way.

The practical consequences are still substantial. Download the weights and run them on your own servers and no data leaves the building, which makes regulated sectors like healthcare and finance viable. Fine-tune on domain data and you get a specialized model. Quantization — trimming precision to run cheaper — requires the original weights, and the community had already posted GGUF and MLX 4-bit builds shortly after release. Above all, weights once released cannot be recalled. A provider raising prices, retiring a model or cutting off a country no longer affects you.

The reverse is also true. Without training data or recipe you cannot rebuild it from scratch, so improvement means waiting for Tencent’s next release. Running 770 billion parameters still requires server-grade hardware, and there is no way to audit what the model was trained on.

The Model Worked on Its Own Development

The most striking claim sits elsewhere. Hy4 preview participated in building itself.

Tencent says the model took part for the first time in automated optimization of training methods, data strategies, evaluation frameworks and low-level operators. It proposed approaches, ran experiments and iterated on results, with the resulting code, logs and feedback feeding back into subsequent rounds. The company calls this an early-stage recursive self-improvement loop.

There is a number attached. The model analyzed bottlenecks in Tencent’s own inference system and carried out several rounds of operator fusion and communication optimization, raising end-to-end throughput 31.8% over baseline, with consistent gains across context lengths and concurrency levels.

Recursive self-improvement has mostly been a topic for US frontier labs. Anthropic has said it expects a similar trajectory in robotics. A Chinese incumbent publishing self-improvement results with a specific figure attached suggests the subject no longer belongs to one side of the industry.

US Labs Close, Chinese Labs Open

Reading this release properly means looking at the strategic split.

US frontier labs keep weights closed. Anthropic, OpenAI and Google all sell their top models through APIs and their own apps rather than publishing weights. The reasons invert cleanly: open the weights and API revenue disappears, competitors inherit hundreds of millions in training spend, and there is no way to prevent fine-tuning that strips out safety measures.

Chinese labs open frontier-class weights instead — DeepSeek, Alibaba’s Qwen, Moonshot AI’s Kimi, Zhipu AI’s GLM, and now Tencent’s Hunyuan. It is the fastest route for a challenger to build an ecosystem. Once developers download and build on a model it becomes a de facto standard, and the money arrives through the products and cloud services layered on top.

The pricing gap is wide. Hy4 preview is priced at $0.834 per million input tokens, $2.501 per million output, and $0.042 for cache hits. Claude Opus 5 runs $5 and $25; Claude Sonnet 5, $2 and $10. Models of different capability cannot be compared on price alone, but where each side wants price to sit is unmistakable.

The tidy story that China opens everything is also wrong. Alibaba has moved to a two-tier approach — mid-tier models opened, the frontier tier kept API-only. Open weights are a positioning tool rather than an ideology, and the top closes once the tool has done its job.

What the Benchmark Comparison Reveals

Tencent’s evidence for the model’s performance makes the market’s structure clearer.

In an internal blind evaluation, 163 experts scored 203 engineering tasks; Hy4 preview averaged 2.99 out of 4.00. The comparisons were GLM-5.3 (2.92) and Kimi K3 (2.94).

Both are Chinese models. GPT-5.5 and Claude Opus 4.8 do not appear. The self-administered nature of the evaluation deserves its own discount, but the more telling point is that Chinese models now benchmark against each other. The framing has shifted from closing a gap with US models toward placing within a domestic ranking.

The capital behind that ecosystem is not small. DeepSeek discussed a $45 billion valuation in its first outside round; Moonshot AI raised $2 billion as its valuation went from $4.3 billion to $20 billion in six months. Alibaba is putting the entirety of an HK$80 billion placement into AI infrastructure.

Distribution is the other axis. Tencent says Hy4 was built in deep co-design with CodeBuddy and WorkBuddy, using training data created with in-house experts across software engineering, gaming, finance and security. Where US labs sell through APIs and chat apps, Tencent plugs the model straight into products hundreds of millions of people already use. Releasing the weights costs it little: the products, the users and the data loop stay with Tencent.

For Korean AI companies the implication is direct. As frontier-class open-weight models keep arriving, the arithmetic between training a foundation model from scratch and specializing published weights shifts. That the party supplying those weights is Chinese remains a separate consideration.

Tencent says the next models in the Hy4 series will follow shortly.