Daily AI CatchupTencent Open-Sources Hy3, a 295B MoE Model, Under Apache 2.0
Tencent shipped the full release of Hy3, a 295B Mixture-of-Experts model that activates only 21B parameters per token yet delivers results competitive with far larger dense models. It's free via OpenRouter for two weeks and fully open on Hugging Face under Apache 2.0 — a permissive license that avoids the regional restrictions that locked out EU, UK, and South Korean users from earlier Chinese model releases.
Why it matters
💻 Developer · 21B active parameters out of 295B total means inference costs closer to a much smaller dense model — worth benchmarking against your current provider while it's free on OpenRouter.
📦 Product · An Apache 2.0 license with no regional restrictions removes a real adoption blocker that's affected prior Chinese open-weight releases — worth reconsidering if you'd previously ruled one out on licensing grounds.
🎨 Design · No direct design impact — this is a backend model release.
📈 Business · Efficient MoE architecture at frontier-adjacent quality, openly licensed, continues to compress the cost advantage that closed frontier labs have historically relied on.
🤔 Just Curious · Tencent released a powerful AI model for free that anyone can download and run themselves, with a license that doesn't restrict which countries can use it — unlike some previous Chinese AI releases.
Try this: Try Hy3 for free via OpenRouter during its two-week free window and compare cost/quality against your current model.
Sources: Tencent's Hy3 on OpenRouter