Tencent’s Open-Source Hunyuan Image 3.0 Jumps to No.1 on LMArena’s Text-to-Image Leaderboard
Want to read in a language you're more familiar with?
Tencent’s open-source Hunyuan Image 3.0 hits No.1 on LMArena’s text-to-image board via blind-vote battles; ranking is preliminary and may shift.
On Oct. 5, 2025, Tencent’s newly open-sourced Hunyuan Image 3.0 has vaulted to the top spot on the LMArena text-to-image leaderboard, edging out high-profile proprietary systems including Google’s “nano-banana” (Gemini 2.5 Flash Image Preview). The ranking is based on LMArena’s human “blind-vote” battles in which users pick preferred outputs without knowing which model generated them.
Hunyuan Image 3.0 is positioned as a native multimodal, autoregressive model aimed at both understanding and generation; Tencent released the model with weights and code under an open-source license on Hugging Face. The model card describes it as a large MoE system and highlights text-to-image quality that “matches or surpasses” leading closed models in internal tests.
LMArena’s latest Text-to-Image Arena page lists “hunyuan-image-3.0” at No.1 as of Oct. 3, 2025 (leaderboard timestamp), while the Arena’s official social account and Tencent Hunyuan’s account also noted the jump. Because the model is new on the platform, its vote count is currently far lower than long-standing entries; the ranking is therefore marked “Preliminary.”
The Hunyuan team has been expanding its open stack in recent weeks, publishing related 3D assets and examples alongside the image model. Developers can download Hunyuan Image 3.0 directly from its Hugging Face repository, and several community demo spaces are already live.
