Alibaba's Z-Image Turbo Tops Open-Weight Rankings on a 16GB GPU
Alibaba's Tongyi-MAI lab released a 6B text-to-image model that renders photorealistic images in under a second on a 16GB GPU, with Apache 2.0 weights.
- Alibaba's Tongyi-MAI released Z-Image-Turbo, a 6B text-to-image model under Apache 2.0.
- Generates 1024x1024 images in 8 steps with sub-second latency on H800 GPUs.
- Runs on consumer 16GB VRAM cards like RTX 3060 and 4090.
- Ranked #1 open-source model on the Artificial Analysis Image Arena, beating FLUX.2 dev and Qwen-Image.
- Uses S3-DiT single-stream architecture plus Decoupled-DMD distillation with reinforcement learning post-training.
- Strong bilingual English and Chinese in-image text rendering; text-to-image only, no edit mode.
Alibaba’s Tongyi-MAI lab has released Z-Image Turbo, a 6-billion-parameter, open-weight diffusion model built for fast text-to-image generation. The Apache 2.0 checkpoint targets photorealistic 1024×1024 output, completes an image in eight model evaluations, and can run on a single GPU with 16GB of VRAM. Tongyi-MAI reports sub-second latency on an H800.
The Comfy-Org mirror displayed more than 9 million Hugging Face downloads at publication. That figure counts file requests, so it represents activity rather than unique users. On the arena leaderboard, Z-Image Turbo ranked as the leading open-weight text-to-image model, ahead of FLUX.2 [dev], HunyuanImage 3.0, and Qwen-Image. The arena measures preferences from blind image comparisons; latency, memory use, and cost require separate benchmarks.
| Specification | Z-Image Turbo |
|---|---|
| Parameters | 6 billion |
| Native target | 1024×1024 pixels |
| Sampling | 8 function evaluations |
| Guidance | No classifier-free guidance pass |
| GPU memory | 16GB reported; actual headroom depends on resolution, precision, and attention backend |
| License | Apache 2.0 |
| Official Turbo input | Text prompts |
Eight evaluations carry the speed
Each number of function evaluations, commonly shortened to NFE, represents one model evaluation during denoising. Many diffusion pipelines use 20 to 50 evaluations to turn random noise into an image. Z-Image Turbo reduces that sequence to eight, cutting both latency and compute.
Classifier-free guidance usually evaluates conditioned and unconditioned versions of the prompt, then combines their predictions to steer the result. Some implementations batch those calculations together, but the additional arithmetic remains. Z-Image Turbo omits that guidance path, removing its compute and memory overhead.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
