If you follow Chinese AI models, you have probably heard of DeepSeek, Qwen (Alibaba), GLM (Zhipu/Z.ai), MiMo (Xiaomi), Kimi (Moonshot), and MiniMax. But until recently, the biggest name in Chinese tech was missing from the open-source AI race: Tencent.
That changed in 2026 when Tencent’s Hunyuan team shipped Hy3, a 295B MoE model with 74.4% SWE-bench Verified, Apache 2.0 licensing, and no geographic restrictions. For the first time, the company behind WeChat, QQ, and some of China’s largest cloud infrastructure is competing openly for developers worldwide.
Here is what you need to know about Tencent as an AI company and why Hy3 matters.
Tencent by the numbers
Tencent is not a startup. It is the largest technology company in China by market capitalization and one of the largest in the world. A few numbers for scale:
- WeChat: 1.3+ billion monthly active users
- QQ: 550+ million monthly active users
- Tencent Cloud: one of the three largest cloud providers in China
- Gaming: owner of Riot Games (League of Legends), stakes in Epic Games, Activision Blizzard
- Revenue: over $85 billion annually
- Employees: 100,000+
For developers, the relevant part is that Tencent has enormous internal demand for AI. With 50+ products using the Hunyuan model family internally (WeChat, QQ, Tencent Meeting, Tencent Docs, gaming, cloud services), the models are battle-tested at scale before they ever reach the public.
Tencent’s AI division: Hunyuan
Tencent’s AI lab is called Hunyuan (literally “primordial” in Chinese). It develops the Hy model family, with Hy3 being the latest. The team sits inside Tencent’s technical infrastructure group and has access to massive compute and internal feedback loops from 50+ products.
Previous Hunyuan models had geographic licensing restrictions that blocked developers in the EU, UK, and South Korea. Hy3 broke that pattern by shipping under Apache 2.0 with no carve-outs, making it the first Tencent model genuinely accessible to developers worldwide.
Why Tencent was the “missing tiger”
The Chinese AI landscape is often described as having “Six Tigers,” the major companies shipping frontier models. Until Hy3, Tencent was the conspicuous absence from the open-source coding model conversation:
| Company | Model | Open weights | Global access |
|---|---|---|---|
| DeepSeek | V4 | Yes | Yes |
| Alibaba/Qwen | 3.7 | Yes | Yes |
| Zhipu/Z.ai | GLM 5.2 | Yes (MIT) | Yes |
| Xiaomi | MiMo V2.5 | Yes (MIT) | Yes |
| Moonshot | Kimi K2.7 | Yes | Yes |
| Tencent | Hy3 | Yes (Apache 2.0) | Yes (new!) |
Now all six have open-weight models accessible globally. The competitive pressure this creates is real: with six strong options, pricing and developer experience become the differentiators. See our full best Chinese AI models 2026 ranking.
What Hy3 brings to the table
The short version: 295B parameters (MoE, 21B active), 256K context, 74.4% SWE-bench Verified, coding on par with DeepSeek V4 Pro while using roughly one-fifth the active parameters. It is fast, cheap to run, and Apache 2.0.
For the full breakdown, read the Tencent Hy3 complete guide. For comparisons, see Hy3 vs DeepSeek V4 and Hy3 vs Qwen 3.7.
What this means for developers
Three things matter practically:
-
Another strong, free, open-weight coding model. If you run models locally or self-host for data residency, Hy3 is now in the mix alongside DeepSeek, Qwen, and GLM. More competition means better options.
-
Apache 2.0 with no geo restrictions. Unlike previous Tencent releases, you can use Hy3 commercially worldwide without worrying about license carve-outs.
-
Battle-tested at Tencent scale. The model was refined with feedback from 50+ internal products between preview and full release. That kind of real-world tuning at scale is hard to replicate in a lab.
The geopolitics angle
As with all Chinese AI models, provenance matters for some organizations. The recent Claude Code steganography finding and Alibaba banning Claude Code show that the US/China AI divide cuts both ways. If your compliance posture requires avoiding Chinese-origin models, Tencent is no exception. If you are comfortable with the landscape, Hy3 is one of the strongest options available.
For the broader supply chain picture, see AI model supply chain risks.
How Hy3 compares to Western models
For context, here is where Hy3 sits relative to the Western flagship models your team might already use:
| Model | SWE-bench | Price per 1M input | License |
|---|---|---|---|
| Claude Sonnet 5 | 63.2% Pro | $2 (intro) | Proprietary |
| Grok 4.5 | 64.7% Pro | $2 | Proprietary |
| Claude Opus 4.8 | 69.2% Pro | $5 | Proprietary |
| Tencent Hy3 | 74.4% Verified | ~$0.14 | Apache 2.0 |
The benchmarks are not perfectly comparable (Verified vs Pro), but the picture is clear: Hy3 is competitive with frontier Western models at a fraction of the cost and with open weights. For the full comparison, see Hy3 vs DeepSeek V4 and Hy3 vs Qwen 3.7.
Frequently asked questions
Is Tencent a new AI company? No. Tencent is China’s largest tech company (WeChat, QQ, gaming). Its Hunyuan AI lab has been developing models for years. Hy3 is its first globally accessible, open-weight frontier coding model.
Is Hy3 free to use? Yes. The weights are Apache 2.0 licensed and available on Hugging Face. It is also available via OpenRouter at approximately $0.14 per million input tokens.
How does Hy3 compare to DeepSeek and Qwen? Hy3 scores 74.4% on SWE-bench Verified and matches DeepSeek V4 Pro on coding benchmarks with only 21B active parameters (roughly one-fifth the compute). See Hy3 vs DeepSeek V4 and Hy3 vs Qwen 3.7.
Can I run Hy3 locally? Yes, with quantization and sufficient hardware. The 21B active parameter count makes it more runnable than full dense models. See how to run Tencent Hy3 locally.
Why did Tencent take so long to release an open model? Previous Tencent models had geographic licensing restrictions. Hy3 is the first under Apache 2.0 with no carve-outs, which required internal policy changes.
The bottom line
Tencent is not a newcomer, it is a giant that finally opened its doors. Hy3 fills the last gap in the Chinese open-source AI landscape and gives developers one more strong, free option for coding work. For the full model details, start with the Hy3 complete guide.