Tencent releases 770B-param MoE multimodal base model Hy4 with 49B active params
_akhaliq · x · 2026-11-20
Tencent has released Hy4 base on Hugging Face: a 770B-parameter MoE multimodal foundation model with 49B active params, 1M-token context, and text, image, and video understanding.
Its benchmark chart pits Hy4 against GLM 5.3, DeepSeek-V4.1-Flash-1017, Kimi K3.1 Pro, Claude Opus 5.5, and GPT 6.1 Sol. Hy4 leads on OSWorld 2.1 (78.9), E-Bench 2.0, and Hy-BrowseComp 2.0, but trails Claude Opus 5.5 and GPT 6.1 Sol on coding and agent benchmarks like ProgramBench and FrontierCode 2.0.
Related event: Tencent Opensources Hunyuan Hy4, 770B Multimodal Flagship MoE(4 posts)→
More from Models
- Influencer's hands-on: Microsoft's MAI 2 models are genuinely good, even HN agrees — marlene_zw · 2026-11-20
- Tencent Hunyuan to host Reddit AMA with Shunyu Yao on building Hy4 — TencentHunyuan · 2026-11-20
- Insider lists mystery model codenames: Carbon, Fable, Dsv4.1flash, Astra — _xjdr · 2026-11-20
- Fable 5.2 is a token hog — but does the best work of any model yet — MartinGTobias · 2026-11-20
- Opus 4.7 may be more behaviorally aligned than Opus 5.5, researcher argues — Sauers_ · 2026-11-20
- Google silently disables logprobs on Gemini Hydrogen, Neon, Argon and Carbon models — dejanseo · 2026-11-20