Z-Image Turbo Locally: Files, VRAM and Settings in ComfyUI

Z-Image-Turbo is the 6-billion-parameter image model from Alibaba’s Tongyi Lab, released on 26 November 2025. Nearly a year later it is still one of the most-used image models in ComfyUI: Comfy-Org’s repackaged files were downloaded about 9.2 million times in the past month, more than any other Comfy-Org repository except MiniMax-H3 and Qwen-Image 2.1.

Three things explain that. It needs only 8 sampling steps, it fits on a 16GB graphics card, and it is released under Apache 2.0, so you can use it commercially without conditions. This guide covers the files, the settings ComfyUI’s templates use, how much memory it needs, and what the rest of the Z-Image family does and does not include yet. BFWR has not run it; memory figures are our calculated estimates, labelled as such.

Quick answer

• Download: about 20.8GB for the standard ComfyUI template, or about 12.1GB for the Int8 version.
• Settings: 8 steps, CFG 1, the res_multistep sampler and simple scheduler, at 1024×1024.
• Memory: about 9GB at 1 megapixel with 8-bit weights and the text encoder offloaded (estimate). Tongyi Lab says it “fits comfortably within 16G VRAM consumer devices”.
• Licence: Apache 2.0, including commercial use.
• Limits: no negative prompts, low variety between seeds, and not meant for fine-tuning. Use the non-distilled Z-Image for those.

The Z-Image family

Tongyi Lab’s model card describes four variants. Two are released:

As of 18 October 2026 the Edit and Omni-Base checkpoints are still listed as “To be released” on both the model card and GitHub, and neither is on Tongyi Lab’s Hugging Face page. The GitHub repository’s last push was on 9 February 2026. So if you need instruction-based editing today, Z-Image is not the answer yet; Qwen-Image 2.1 and FLUX.3 Image Klein both edit in one checkpoint, under more restrictive licences.

All the models share one architecture, which Tongyi Lab calls a Scalable Single-Stream DiT: text, visual semantic tokens and image tokens go through the same transformer as a single sequence. The text encoder is Qwen3 4B and the image decoder is the FLUX.1 VAE, which ComfyUI’s example page notes you may already have.

Turbo or the base model?

The base model’s card has a direct comparison:

In practice: use Turbo to make images quickly, and expect similar-looking results when you change only the seed, because distillation narrows the output. Use the base model when you want variety, need a negative prompt to remove something, or plan to train a LoRA; its card calls it “a good base for LoRA training”. The base model’s recommended settings are a guidance scale of 3.0 to 5.0 and 28 to 50 steps, at 512×512 up to 2048×2048 in total pixel area.

Running it in ComfyUI

ComfyUI supports Z-Image natively. Its template index lists ComfyUI 0.11.0 as the minimum for the Turbo templates and 0.23.0 for the base model, so any current install has both. The templates:

• Z-Image-Turbo: Text to Image, with the BF16 model
• Z-Image-Turbo Int8: Text to Image, a smaller 8-bit version added in July 2026
• Z-Image: Text to Image and an Int8 version, for the base model
• Z-Image-Turbo Fun Union ControlNet, which ComfyUI describes as supporting “Canny, HED, Depth, Pose, and MLSD”
• Image Upscale: Z-Image-Turbo 2K, in the Image Tools group

The Turbo template’s KSampler uses 8 steps, CFG 1, the res_multistep sampler and the simple scheduler, at 1024×1024, with a ModelSamplingAuraFlow shift of 3. The base-model template uses 25 steps and CFG 4. The text encoder loads through Load CLIP with the type set to lumina2. If you use diffusers instead, Tongyi’s example sets num_inference_steps=9, which its comment says “actually results in 8 DiT forwards”, and a guidance scale of 0.

Files and download size

None of the files are gated. Sizes from Comfy-Org’s listing:

ComfyUI’s template index gives 20.8GB for the standard Turbo template and 12.1GB for the Int8 one. The base model’s files are the same sizes, with z_image_bf16.safetensors in place of the Turbo file. The NVFP4 file uses NVIDIA’s 4-bit format for its RTX 50-series cards. Note that the Qwen3 4B encoder here is the plain text model, not the Qwen3-VL 4B used by Krea 2 and FLUX.3 Image Klein, so the files are not interchangeable.

How much memory it needs

These are estimates from the formula behind our image generation VRAM calculator: weights at each precision plus an activation allowance per megapixel. Tongyi publishes no activation figure, so the allowance is borrowed from FLUX.1. Treat these as planning figures with headroom.

That lines up with Tongyi’s 16GB claim: the BF16 model fits on a 16GB card at 1 megapixel once the text encoder is offloaded, which ComfyUI does automatically when memory is short. With the Int8 files, a 12GB card has room to spare at 1 megapixel. Our Z-Image-Turbo VRAM page shows every precision and offload mode.

For comparison, on the same formula at 1 megapixel with 8-bit weights and the encoder offloaded: FLUX.2 [klein] 4B needs about 6.8GB, MOSS-Image 8.1GB, FLUX.3 Image Klein 8.7GB, Qwen-Image 2.1 10.8GB and Krea 2 16.2GB. Z-Image-Turbo sits in the middle of this year’s models on memory.

The licence

The weights, the base model and the GitHub code are all Apache 2.0, as are Comfy-Org’s repackaged files. There is no revenue threshold, no content-filter clause and no naming rule. That sets Z-Image apart from most of 2026’s popular image models:

Our guides to FLUX.3 Image Klein, Krea 2 and Qwen-Image 2.1 cover those terms. As always, a community LoRA or merge built on Z-Image carries whatever licence its author chose, so check each one.

LoRAs and ControlNet

Because Turbo is distilled, Tongyi marks it as not fine-tunable and points to the base model for training. LoRAs built for Z-Image are specific to this architecture and will not load on FLUX or Qwen-Image models; our guide to matching LoRAs to their base model explains why. For pose and structure control, the Fun Union ControlNet template adds one model patch, Z-Image-Turbo-Fun-Controlnet-Union.safetensors, in ComfyUI’s model_patches folder. Our ControlNet guide covers how control models work in general.

FAQ

Is Z-Image-Turbo free for commercial use?

Yes. It is Apache 2.0, with no revenue limit.

Can Z-Image-Turbo run on 8GB of VRAM?

Possibly at 4-bit with the text encoder offloaded, which we estimate at about 6.7GB at 1024×1024. That leaves little headroom. 12GB is a more comfortable minimum, and 16GB runs the full BF16 model.

Why does Z-Image-Turbo ignore my negative prompt?

Turbo runs without classifier-free guidance, and Tongyi’s card lists negative prompting as unsupported for it. Use the base Z-Image model if you need one.

Has Z-Image-Edit been released?

Not as of 18 October 2026. Tongyi Lab still lists it as “To be released”.

Related guides and tools

• MOSS-Image in ComfyUI, another Apache 2.0 model from 2026
• ComfyUI templates: local versus API
• Best local image generation models in 2026, every model compared on memory and licence
• ComfyUI hub and image model directory
• Stable Diffusion Prompts
• AI Upscale Models in 2026

Sources

• Tongyi-MAI/Z-Image-Turbo model card and Z-Image model card
• Tongyi-MAI/Z-Image on GitHub (model zoo and news)
• Comfy-Org/z_image_turbo repackaged files
• ComfyUI docs: Z-Image-Turbo and workflow templates

Updated 18 October 2026. File sizes, download counts, template settings and release status were read on that date. Memory figures are estimates, not measurements.