Hy4
AvailableTencent's fourth-generation flagship LLM (Hy series, formerly Hunyuan), released Nov 4 2026 as the final version after Hy4 preview (Aug 28). It is a natively multimodal MoE (text, image and video understanding; text output) with a 770B-parameter language backbone and 49B active per token (plus a 1.23B ViT), 78 layers (1 dense, 77 MoE with 256 routed experts, top-8 plus 1 shared), Gated DeepSeek Sparse Attention with IndexCache, iHC residuals, a built-in 3.3B DFly2 speculative-decoding drafter, and a 1,048,576-token context. Tencent says it significantly scaled pre- and post-training, adding visual understanding for the first time plus gains in instruction following, long-context comprehension and long-horizon agentic tasks. Available via Tencent Cloud TokenHub, WorkBuddy, CodeBuddy, Yuanbao, WeCom and ima, with API access through OpenRouter and other third parties, at about $0.90 / $2.70 per 1M input/output tokens ($0.045 cached input). Open weights followed on Nov 20 2026 on Hugging Face, ModelScope and GitCode (Hy4, Hy4-FP8 and the pre-trained Hy4 base, ungated) under the Tencent Hy License 1.0, an Apache-2.0-derived license where large Model-as-a-Service businesses need separate written authorization. Unlike Hy4 preview, which was Apache 2.0. Benchmarks vendor-reported.
Specifications
- License
- Open weights · Tencent Hy License 1.0
- Weights
- Downloadable
- Architecture
- Mixture-of-Experts
- Parameters
- 770B · 49B active
- Context window
- 1.0M tokens
- Max output
- —
- Knowledge cutoff
- —
- Price (in / out, $/M)
- $0.9 / $2.7
- Modalities
- TextVisionCode
Benchmarks
No benchmark scores recorded yet. Spotted some? Submit a correction.
Vendor-reported figures are claims until independently verified. See methodology.