Hy3 preview is a sparse Mixture-of-Experts model developed by the Tencent Hy Team, featuring 295 billion total parameters with only 21 billion activated per inference alongside 3.8 billion parameters in multi-token prediction layers. This selective routing enables substantial capacity while keeping computational requirements manageable. The model introduces a hybrid fast-and-slow-thinking architecture, blending rapid response generation with deeper deliberation pathways. It was the first model trained on Tencent's rebuilt infrastructure, and third-party reports indicate significant improvements in complex reasoning, instruction following, in-context learning, coding, and agent-oriented tasks compared to earlier Tencent models.
The model supports a 256,000-token context window suited for extensive document analysis and long-form reasoning workloads. It was open-sourced under the Apache 2.0 license in July 2026, building on the preview lineage launched in April 2026, making it accessible for self-hosting and customization. Within its first week on OpenRouter, the model processed 6.13 trillion tokens and reached the top of the platform's weekly usage rankings. Distribution paths include Tencent Cloud's TokenHub for paid API access, alongside free availability on WorkBuddy through late August 2026. For developers, the model fits well in code generation assistants, agent frameworks, and retrieval-augmented generation pipelines that benefit from extended context and tool-based workflows.
