Hy3 is Tencent's open-source large language model built around a Mixture-of-Experts architecture, pairing 295 billion total parameters with 21 billion active per query and a routing scheme that activates the top 8 of 192 specialised experts. The design keeps inference demands manageable while preserving access to a broad skill set, and Tencent released the full version after running an internal preview across 50 product teams, an iteration cycle that reportedly drove hallucinations down from 12.5% to 5.4% and cut commonsense reasoning errors in half. Weights are published on Hugging Face under the tencent organisation, and the Apache 2.0 license makes the model straightforward to self-host or fine-tune for commercial deployments.

The model targets long-horizon, agent-driven work. It advertises a 256K-token context window suited to enterprise documents, multi-turn constraint tracking, and stable tool calling across different agent frameworks. Configurable reasoning effort lets callers choose a fast no-think mode by default or escalate to low and high chain-of-thought settings for complex math, coding, and multi-step planning, and Tencent positions Hy3 for grounded behaviour in domains such as financial analysis, document processing, game development, and frontend design. Through this free-tier listing, teams can route requests to that capability set at zero listed cost, making it a practical option for evaluating large-scale agentic systems before committing to heavier infrastructure.