A video editor in Guangzhou types one sentence and receives a ten-second clip. A game studio generates a 3D asset from a text prompt. A developer in Berlin clones a model from GitHub that, a year earlier, lived only inside one Chinese internet company.
That company is Tencent (腾讯), and the model family is Hunyuan (混元). Over roughly twelve months, Tencent went from a cautious, mostly closed AI player to one of the most prolific open-source publishers in China. The move is less about altruism than a simple cloud-era equation: give the model away, keep the bill for compute.
The largest open MoE, then 3D
The opening shot came on November 5, 2024. Tencent released Hunyuan-Large, a mixture-of-experts model with 389 billion total parameters and 52 billion active at any time, trained on about 7 trillion tokens and able to handle a 256,000-token context. At launch it was the largest open-source MoE model available, and it was free to use commercially.
On the same day Tencent shipped Hunyuan3D-1.0, described as the first open-source model that could turn both text and images into 3D assets. For game and industrial designers, that meant generating a chair, a building, or a tool from a prompt without a modeling team.
Then video
A month later, on December 3, 2024, Tencent open-sourced HunyuanVideo, a roughly 13-billion-parameter video generation model built on a DiT (diffusion transformer) architecture and released under the permissive Apache 2.0 license. At the time it was the largest open video generation model, and Tencent said it outperformed several closed competitors on text-video alignment and motion quality.
The pattern is deliberate. Tencent did not open one flagship model; it opened a stack — language, 3D, video — each one a building block for the creative and enterprise workloads that run on its cloud.
A fast model you can actually talk to
The consumer-facing side arrived in March 2025. On March 1, Tencent added Hunyuan Turbo S to its Yuanbao (元宝) assistant app. Where deep-reasoning models "think before answering," Turbo S is built to reply instantly — Tencent reported output speed roughly doubled and first-token latency cut by 44%. It is the model you get when you want an answer now, not a chain-of-thought essay.
Through Tencent Cloud's API, Turbo S is priced at 0.8 yuan (≈ US$0.11 / HK$0.9) per million input tokens and 2 yuan (≈ US$0.28 / HK$2.2) per million output tokens — a direct answer to the price pressure DeepSeek put on the whole Chinese market in early 2025. Tencent also said Turbo S is the first industrial-scale application of a Hybrid-Mamba-Transformer fusion to a super-large MoE model, which is what lets it cut memory and compute cost.
The discipline behind the generosity
Tencent's open-source choices are not random. Kang Zhanhui (康战辉), the company's machine-learning platform director and a lead on the Hunyuan language model, put it plainly: Tencent does not open models for the sake of opening them; it open-sources models already hardened inside its own products. The logic is that a model carrying real production scars is more useful to the outside world than a research curiosity.
That is why Hunyuan's public releases keep mapping onto Tencent's own businesses — content, gaming, advertising, cloud. The model is the demo; the cloud bill is the business.
Why a gaming-and-social giant plays this game
Tencent is, first, an infrastructure seller. Its incentive is not to win the chatbot popularity contest but to make Hunyuan the default engine inside as many products as possible, then charge for the GPUs and NPUs those products consume. Open weights lower the barrier for a startup to build on Tencent Cloud instead of a rival's. In that sense Tencent's strategy resembles Microsoft's with open models and Azure, or Meta's with Llama and its ad-and-service ecosystem: the model is a customer-acquisition cost.
Yuanbao (元宝), meanwhile, is the consumer face — an assistant stitched into WeChat (微信) search and the broader Tencent app graph. It lets ordinary users touch Hunyuan without knowing the weights exist.
The company Tencent is really competing with
The obvious comparison is Alibaba's Qwen, the other Chinese cloud giant using open models to pull developers toward its cloud. Tencent's differentiator is media: video and 3D are where it has consumer and gaming DNA. The quiet contest between the two is less about which model is smartest and more about which ecosystem a builder lands in first.
What readers can do now
- If you build with generative media, test HunyuanVideo and Hunyuan3D-1.0 — both are open and free to prototype, which removes the usual licensing friction for commercial pilots.
- If you run inference at scale, benchmark Turbo S's API price (0.8 / 2 yuan per million tokens) against your current provider; the gap may pay for a migration.
- If you track the China AI market, read Tencent's cadence (language → 3D → video → fast model) as a template for how a cloud vendor turns research into a product ladder.
Honest limitations
Facts here come from Tencent's own developer channels, the Hunyuan GitHub repository and the arXiv paper (2412.03603), Chinese outlets including Huanqiu, Beijing News, and Sina, plus a Tencent Cloud developer article. The "largest open-source MoE / video model at the time" claims reflect Tencent's launch statements and contemporaneous reporting, not an independent benchmark I ran. The benchmark comparisons Tencent cites for HunyuanVideo and Turbo S are vendor-reported, and the Mamba-architecture claim is Tencent's own characterization. I have not deployed these models myself, so performance and cost figures should be treated as vendor claims to be verified by your own workload tests.