AI news ·
Tencent open sources Hy4 model, completing full release of Hunyuan LLM series
Tencent open-sourced its 770-billion-parameter Hunyuan Hy4 model under Apache 2.0, giving developers unrestricted commercial use and a 1-million-token context window. A cost-optimized Hy4 flash variant with 224B parameters offers roughly 20x lower per-task cost for scaled deployment.

Tencent's Hunyuan team open-sourced the Hy4 large language model on November 20, 2026, releasing model weights across GitHub, Hugging Face, ModelScope, and GitCode. The move completes the full open-sourcing of the Hunyuan LLM series, giving developers unrestricted access to a 770-billion-parameter Mixture of Experts model with a 1-million-token context window.
Hy4 was first released in its production form on November 4. The model uses a MoE architecture with 770B total parameters and activates 49B parameters per token. Alongside the flagship model, Tencent also released Hy4 flash, a cost-optimized variant with 224B total parameters, 64B Engram parameters, and 12B activated parameters per token. Hy4 flash supports multimodal input and the same 1-million-token context length, with API pricing significantly lower than the full Hy4 model.
Licensing and safety disclosures
Both Hy4 and Hy4 preview ship under the Apache 2.0 license, which permits commercial use and modification. Tencent published a System Card alongside the release, detailing safety evaluations across cybersecurity and biosafety dimensions. The training data included only defensive cybersecurity content - proof-of-concept generation and vulnerability fixes - while excluding offensive material such as exploit development, reverse engineering, and penetration testing. Audits by Tencent's Xuanwu and Keen Labs found no evidence of offensive capability risk.
What the full Hunyuan lineup covers
With Hy4's release, the Hunyuan LLM portfolio is now fully open, spanning Hy3 preview, Hy3, Hy4 preview, Hy4, and Hy4 flash. The range gives teams flexibility to choose between maximum capability and cost efficiency depending on the workload. Model weights can be downloaded directly from any of the four hosting platforms, and a GGUF quantized version is available through a community contributor on Hugging Face.
Why this matters for product and engineering teams
For product developers and engineers, the Apache 2.0 license removes the usual friction around commercial deployment. You can fine-tune Hy4 on proprietary data, embed it in production applications, and distribute derivative models without negotiating separate terms. The 1-million-token context window also changes what's practical to build - full codebases, lengthy documentation sets, and multi-hour transcripts can be processed in a single pass. Teams weighing cost against performance can prototype on Hy4 and switch to Hy4 flash for scaled deployment, with the smaller model offering roughly a 20x cost reduction per task according to Tencent's earlier disclosures. Professionals looking to build on these models can explore Generative AI Courses or the AI Engineering Courses learning path for hands-on implementation guidance.