Tencent Open-Sources Hunyuan Hy4 Preview: 770B Total Parameters, 49B Active, 1M+ Context

Tencent released Hunyuan Hy4 preview under open weights with API access via TokenHub and OpenRouter. The model expands prior Hunyuan scale in parameters, context length, and training volume while incorporating domain data from software engineering, finance, and game development. It integrates directly into WorkBuddy and CodeBuddy for productivity workflows and ships with a two-week free access window.

Internal blind evaluation across 163 experts and 203 engineering tasks produced 2.99/4.00, edging GLM-5.3 at 2.92 and Kimi K3 at 2.94. Separate inference optimizations delivered a measured 31.8% end-to-end throughput increase through operator fusion and communication tuning. The model also executed automated iterations on its own training pipeline and low-level kernels, closing an early recursive loop.

Hy4 preview sits within China's pattern of releasing large open MoE systems that match or exceed Western closed models on targeted productivity benchmarks. The 49B active parameter design reduces inference cost relative to dense equivalents while the 1M context directly addresses long-document and multi-file code tasks that remain bottlenecks for most open models. Continued product co-design with Tencent applications provides usage data that closed Western labs lack.

Operational deployment will likely concentrate on enterprise API usage and continued internal self-optimization loops rather than broad consumer chat. Next measurable signals are external benchmark releases and whether the 31.8% throughput gain holds under third-party inference stacks.