Quick Answer: Tencent’s Hy3 is a 295-billion-parameter open-source Mixture-of-Experts AI model released on April 23, 2026. It activates 22 billion parameters per token, supports a 256K context window and scored 74.4% on SWE-bench Verified. Placing it among the strongest open-source large language models available right now.
Tencent dropped Hy3 last week and developers are already stress-testing it. It’s not from OpenAI. It’s not from Google. And within days of going public, it landed in the middle of one of the most talked-about controversies in AI this year. Having tracked Chinese AI model development closely here at TechGlimmer, this release stands out not just for the benchmark numbers, but for the questions it raises about how frontier AI actually gets built.
What Is the Tencent Hy3 AI Model?
Hy3 is Tencent’s latest large language model, built on a Mixture-of-Experts (MoE) architecture with 295 billion total parameters. But it only activates 22 billion of them at a time during inference.
Think of it like a hospital. Not every specialist shows up for every patient — only the right ones do. Hy3 works the same way: it routes each input to the most relevant slice of its network. You get cheaper inference without the usual performance hit.
Here are the core specs:
- Total parameters: 295 billion
- Active parameters per token: 22 billion
- Context window: 256K tokens — enough to process a full-length novel in a single prompt
- Availability: Open-source on Hugging Face and Tencent Cloud
Built in Under Three Months
Tencent built and open-sourced Hy3 in under three months. From a cold infrastructure start in February 2026 to public release on April 23, 2026.
Building a frontier model in under three months is the kind of timeline that makes other AI labs uncomfortable. The team overhauled both the pre-training pipeline and the reinforcement learning infrastructure from scratch. Leading the charge was Yao Shunyu, a former OpenAI research scientist Tencent recruited specifically to push their AI division harder and faster.
The Benchmark Numbers
On SWE-bench Verified — one of the most trusted software engineering benchmarks in AI Hy3 scored 74.4%. Its predecessor Hy2 scored 53%. That’s not a polish-and-ship update. That’s a generational jump inside a single release cycle.
| Benchmark | Hy3 Score | Hy2 Score |
|---|---|---|
| SWE-bench Verified | 74.4% | 53% |
| BrowseComp | 67.1% | — |
For a free, open-source model, these numbers are genuinely hard to argue with. It still trails leading closed models from OpenAI and Google in certain reasoning tasks but the gap is closing fast.
The Part That Got Everyone Talking
Shortly after launch, The Information reported that Tencent employees used Anthropic’s Claude chatbot to help fine-tune Hy3. This process known as model distillation involves using outputs from one AI system to train or sharpen another.
And the timing here is genuinely awkward. Back in February 2026, Anthropic publicly accused three Chinese AI companies DeepSeek, Moonshot AI and MiniMax — of running what it called industrial-scale distillation campaigns against Claude. Roughly 24,000 fraudulent accounts. Over 16 million interactions. All to extract Claude’s knowledge and funnel it into competing models.
Anthropic’s terms of service explicitly ban competitors from using Claude outputs to train rival AI. Its services are also officially restricted in China. Whether Tencent’s use crossed a contractual line hasn’t been confirmed by either side — but the question is out there now, and it won’t go away quietly.
Zoom Out and It Gets Bigger
This isn’t a one-off incident — it’s a pattern and nobody’s figured out how to stop it yet.
Open-source AI is supposed to encourage collaboration and lower barriers to entry. But when a major Chinese tech company uses an American AI safety company’s model to train a competitor intentionally or not. It stops being just a product story. It becomes a geopolitical one. Washington and Beijing are both drawing harder lines around technology transfer and AI sits right in the middle of that fight.
For developers, Hy3 is worth testing today. It’s free, powerful and already gaining serious traction on Hugging Face. But we’ll keep a close eye on whether Anthropic pushes back because that response, if it comes could be messier than the model launch itself.
You might be interested in following article
What Are AI Connectors? The Bridge Between AI and Your Favorite Tools
