SourcesTencent Hunyuan
publisher · 4 sources
Tencent Hunyuan
4 sources read from this publisher, newest first, with the methods each one describes.
tencent/Hy3 · Hugging Face
first party release · model card · 2026-07-06 · 11 methods
tencent/Hy4-preview · Hugging Face
first party release · model card · 2025-12-02 · 15 methods
Mixture of expertsMulti-Token PredictionFP8Speculative decodingShared expertTop-k expert routingChain-of-thought reasoningNEXTN speculative decodingBlind expert evaluation of engineering tasksGated DeepSeek Sparse AttentionIdentity Hyper-ConnectionsIndexCacheAngelSlimExpert co-created training datavLLM or SGLang serving
reasoning_effort: "no_think" (default, direct response), "low", "high" (deep chain-of-thought)
first party release · code repo · 11 methods
Tencent-Hunyuan/Hy4-preview
first party release · code repo · 12 methods
Mixture of expertsMTP speculative decodingFP8Shared expertBlind expert evaluation of engineering tasksGated DeepSeek Sparse AttentionIdentity Hyper-ConnectionsIndexCacheCo-designing the model with downstream productsEarly release with feedback-driven iterationLarge-model quantization and speculative samplingTraining-data construction around expert work products