Tencent has released Hy3, a Mixture-of-Experts model with 295 billion total parameters (21 billion active) under Apache 2.0 license. The model was shaped by feedback from over 50 internal products following an April preview, and according to the company, it outperforms similarly-sized models and approaches much larger open-source flagships. It features a 256K token context and weighs 598 GB in full form, 300 GB when quantized to FP8.

Why it matters to you. It’s one of the first open options at this scale from a Chinese player, with licensing that lets you use, modify, and deploy it to production without restrictions. If you’re in an environment where self-hosting is mandatory (sensitive data, regulated sector, zero dependence on external APIs), Hy3 enters the shortlist of alternatives to Llama or Mistral Large. The active size (21B) makes it more manageable than it appears: it runs on reasonable infrastructure, no datacenter required.

If you want to try it. Hy3 is available on Hugging Face and free on OpenRouter until July 21st. You can test it immediately via API without downloading anything, to see whether its behavior on real tasks (long summarization, technical translation, structured generation) justifies setting it up to run locally. The quantized model weight (300 GB) still requires dedicated GPUs or cloud instances with sufficient memory.