Z-Image Turbo

• 🆕 Distilled, speed-optimized variant of the 6B-parameter Z-Image model
• 📏 Generates 1024×1024 photorealistic images in sub-second time
• ⚡ Uses an 8-step sampler (8 NFEs) instead of dozens of diffusion steps
• 🔧 Single-stream diffusion transformer with guidance baked into the weights
• 🌐 Bilingual English–Chinese on-image text rendering
• 🏢 Released by Alibaba's Tongyi Lab (Tongyi-MAI) under Apache 2.0
• 👁️ Runs within 16GB VRAM on consumer cards like RTX 4090
• 🎯 Sub-second latency on enterprise H800 GPUs

Z.ai

Z.ai, formally Knowledge Atlas Technology Joint Stock Co., Ltd., is a Chinese technology company specializing in artificial intelligence. Previously known internationally as Zhipu AI, the company rebranded to Z.ai in 2025. Its core focus is the GLM family of…

Z-Image Turbo is a 6-billion-parameter text-to-image model built on a single-stream diffusion transformer, published on Hugging Face by Alibaba's Tongyi Lab (the Tongyi-MAI team) under the permissive Apache 2.0 license. Its central design goal is practical deployment: per the official model card, the model fits within 16GB of VRAM on consumer graphics cards like the RTX 4090 while still producing photorealistic output, and it emphasizes bilingual English–Chinese text rendering directly within images.

The key improvement over its same-family base model is distillation. According to the official model card, Z-Image Turbo is a distilled version of Z-Image that reaches comparable quality with only 8 NFEs (Number of Function Evaluations), where standard diffusion models often need dozens of steps. In practice this means running roughly eight forward passes at 1024×1024 with classifier-free guidance disabled, since guidance is baked into the distilled weights.

This few-step approach yields sub-second inference latency on enterprise-grade H800 GPUs while remaining usable on consumer hardware. The vendor reports evaluating the model via an Elo-based human preference comparison on Alibaba AI Arena.

For builders, the model targets interactive and high-throughput use cases—real-time web features, dashboards, and batch backends—rather than only offline jobs, helped by an integrated prompt enhancer that refines inputs for stronger instruction adherence.