Tencent's Free Hy3 Is a 295B Model That Rivals Trillion-Scale Flagships

Tencent has released Hy3, the latest entry in its Hunyuan model family, under the Apache 2.0 licence. The model uses a Mixture-of-Experts architecture with 295 billion total parameters, but only 21 billion are active for any given token. This design gives Hy3 a 256,000 token context window while keeping inference costs far below what a dense trillion-parameter model would demand, allowing it to compete with much larger flagship systems without the associated running costs. Tencent has also made significant progress on reliability. Hallucination rates have dropped from 12.5% in the April preview version to 5.4% in this release, a substantial improvement for a model of this scale. Alongside the open-source release, Tencent is offering a free two-week API access window on OpenRouter, giving developers and researchers a no-cost route to test the model before deciding whether to integrate it into paid workflows. Hy3 isn't just a research release, it's already deployed in production. Tencent has integrated it into WeChat's assistant features, putting the model in front of one of the largest user bases in the world, and it's also running inside Tencent's own game studios, suggesting internal confidence in its performance and cost profile for real-world applications. The release fits a broader pattern of Chinese AI labs pushing efficient, open-weight models that challenge the assumption that frontier performance requires trillion-parameter scale. By combining a large total parameter count with sparse activation, Tencent is positioning Hy3 as a model that offers flagship-level capability without the infrastructure burden, and doing so under a permissive licence that allows commercial use and modification.