tencent/Hy3
Simon Willison's Weblog Simon Willison ● Covered by 2 sources
Tencent just dropped Hy3, a huge open-source AI model that's free to try right now. It punches above its weight, matching models 2-5x its size, and you can test-drive it before July 21st.
Based on reporting by Simon Willison's Weblog, Simon Willison — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Tencent's Hy Team has released Hy3, and the numbers alone make it worth a second look. This is a 295-billion-parameter Mixture-of-Experts model, though only 21 billion of those parameters activate at any given time, plus another 3.8 billion tucked into an MTP layer. That's the kind of architecture that lets a model punch well above its file size, and Tencent is claiming Hy3 rivals flagship open-source models carrying two to five times as many active parameters.
This isn't a first swing at the idea. Tencent shipped a Hy3 Preview back in late April, and apparently it made enough of an impression that developers noticed before the official announcement caught up. Max Woolf wrote about that earlier build in May, when it was quietly climbing to the top of OpenRouter's model rankings without much fanfare. Tencent says it took feedback from over 50 products built on that preview and used it to scale up post-training with better-quality data, which is a fairly deliberate way of building a model in public before the real launch.
The practical footprint is substantial. The full model weighs in at 598GB on Hugging Face, while an FP8 quantized version cuts that roughly in half to 300GB. Context length runs to 256K tokens, which puts it in the same range as other serious long-context models being released this year. And crucially, it's Apache 2.0 licensed, so anyone can self-host or fine-tune it without the usual licensing headaches that come with some Chinese model releases.
Simon Willison ran his now-standard test, asking Hy3 to generate an SVG of a pelican riding a bicycle, available for free on OpenRouter until July 21st. The result was noticeably better than what the April preview produced. That earlier version, interestingly, came with a button to change the pelican's color, an oddball feature Willison hadn't seen from any other model before or since. Today's version didn't include that flourish, but the bicycle rendering itself was cleaner and more coherent, suggesting the extra post-training round genuinely improved the model's grasp of spatial reasoning, not just benchmark scores.
My take — AI-written commentary, not fact-checked reporting
Another week, another enormous open-weight model out of China outperforming stuff with far more marketing budget behind it, and I think Western AI labs keep underestimating how fast this pipeline of preview-then-refine releases is closing the gap. Apache 2.0 licensing on a 295B MoE model is the real story here, not the pelican drawings, because it means anyone with the hardware can actually build on this rather than just admire it from behind an API paywall.”
Read more about this at: Simon Willison's Weblog