Tencent open-sourced Hunyuan-DiT, a bilingual Chinese-English text-to-image Diffusion Transformer pairing dual CLIP and multilingual T5 text encoders with a MLLM conversational prompt-refinement loop.
Tencent Hunyuan
Also known as: Hunyuan · Hunyuan-Large · Hunyuan-DiT · Hunyuan3D · Hy4 Flash
Plain English
Tencent's foundation model division—builder of the open-weight HunyuanVideo video generator, Hunyuan3D 3D asset generator, Hunyuan-DiT image model, and the Hunyuan-Large and Hy4 Flash Mixture-of-Experts language models.
Technical Overview
Tencent Hunyuan develops open-weight and enterprise foundation models across modalities: bilingual image DiTs (Hunyuan-DiT), 13B flow-matching HunyuanVideo, two-stage shape-and-texture 3D generation (Hunyuan3D 2.0), and large sparse MoE LLMs (Hunyuan-Large 389B/A52B and Hy4 Flash 220B/A12B with hybrid attention and LatentMoE).
Timeline (4)
2024
Paper- Paper
Tencent released Hunyuan-Large (389B total / 52B active parameters), at the time the largest open-weight Transformer MoE model, combining shared+specialized expert routing, KV cache compression, and 256K context support—followed a month later by HunyuanVideo.
2025
PaperTencent open-sourced Hunyuan3D 2.0, decoupling high-resolution 3D asset creation into a flow-matching geometry diffusion transformer (
Hunyuan3D-DiT) and a mesh-conditioned multi-view texture diffusion model (Hunyuan3D-Paint) with ComfyUI and Blender plugins.2026
ReleaseTencent released Hy4 Flash, a ~220B–224B total / 12B active multimodal MoE with a 1M-token context window that combines hybrid attention, LatentMoE, and Engram lookup structures to shrink KV cache footprint by 6x.
See also (10)
Open full graphTencent Hunyuan