Tencent Releases Hy3: 295B MoE, 21B Active, 256K Context, Apache 2.0

Tencent's Hy team released Hy3, a 295B-parameter mixture-of-experts model with 21B active parameters and a 3.8B MTP layer. The full model is 598GB on HuggingFace, with a 300GB FP8 quantized version, 256K context length, and Apache 2.0 license. After the Hy3 Preview in late April, the team collected feedback from over 50 products and expanded post-training with higher-quality data. The final version performs best among similar-sized models and rivals open-source flagships with 2-5x more parameters, showing significant gains across various product and productivity tasks. Currently, Hy3 is free on OpenRouter until July 21. Developers using Claude API or self-built inference services can test this lightweight yet powerful model on OpenRouter. It is especially suitable for long-context (256K) scenarios with low hardware requirements—the FP8 quantized version is only 300GB, much smaller than full models of similar capability. Teams running local open-source models that previously required over 600GB will see a significant reduction in resource usage with Hy3.