Z-Image Turbo, 4 images at 1024×1024
- RTX PRO 6000 96 GB, $2.42/h
- 6.5 s · $0.004
- RTX 6000 Ada 48 GB, $1.37/h
- 17.7 s · $0.007
2.7 times faster, and 35% less for 4 images.
Z-Image
Z-Image (Turbo) needs 12.7 GB of VRAM in ComfyUI, at the diffusion stage of its own template, by file sizes read 3 November 2026. The smallest card that loads it with room to work is the RTX 5080 (16 GB), and the template downloads 20.7 GB.
- Made by
- Tongyi-MAI (Alibaba)
- Licence
- Apache 2.0
- Released
- November 2025
- Sizes read
- 3 November 2026
Needs
12.7 GB
At the diffusion stage, with the BF16 file ComfyUI's template loads. The template downloads 20.7 GB.
Computed 3 November 2026
Smallest card
16 GB
The RTX 5080 loads it with 3.3 GB left for the work.
With the smaller INT8 ConvRot file, the RTX 5070 loads it.
Computed 3 November 2026
Runs on Comfy Cloud?
Yes
ComfyUI's template for it uses built-in nodes only, and its largest stage, 12.7 GB, fits the 96 GB RTX PRO 6000.
Read 3 November 2026
Cheapest run we know
$0.004
Our run on a rented RTX PRO 6000: Z-Image Turbo, 4 images at 1024×1024, in 6.5 s.
Measured 20 October 2026
Largest stage from exact file sizes, with ComfyUI's own template; LoRAs are not counted. Working memory for the generation comes on top. Open it in the ComfyUI checker
Needs
12.7 GB
At the diffusion stage, with the BF16 file ComfyUI's template loads. The template downloads 20.7 GB.
Computed 3 November 2026
Largest stage from exact file sizes, with ComfyUI's own template; LoRAs are not counted. Working memory for the generation comes on top. Open it in the ComfyUI checker
Which file on your card
For each memory size, the best file for that card: the highest-precision diffusion file that loads with room to work (one that only just fits when nothing else does), and what is left for working memory while it samples. On a larger card that can be above the file ComfyUI's template loads, which the answer above is about. "Streams" means even the smallest file is larger than the card: ComfyUI still runs it from system RAM, more slowly. FP4 files go only to the Blackwell cards that run them natively.
Turbo
ComfyUI's template for Turbo: BF16 diffusion model, BF16 text encoder. Stages: text encoder 8.0 GB, diffusion 12.7 GB. Download 20.7 GB.
Computed 3 November 2026Base
ComfyUI's template for Base: BF16 diffusion model, BF16 text encoder. Stages: text encoder 8.0 GB, diffusion 12.7 GB. Download 20.7 GB.
Computed 3 November 2026Check another card, or pick a file by hand, in the ComfyUI VRAM checker. Built your own workflow around it, with LoRAs or custom nodes? Drop it into the ComfyUI workflow checker.
Where to run it
Turbo
Your own GPU
16 GB or more
Loads ComfyUI's template with room to work. 12 GB loads the smaller INT8 ConvRot file.
Computed 3 November 2026
Comfy Cloud
$4.54 an hour of generation
A 96 GB RTX PRO 6000, billed only while a workflow runs. Plans from $20 a month.
Read 3 November 2026
A GPU by the hour
$0.17 an hour
An RTX 3090 holds it with room to work, on SaladCloud, or $0.50 on RunPod, which runs its own data centres. Any node; it bills while it idles.
Read 19 October 2026
On-demand prices before tax, read 19 October to 3 November 2026.
Base
Your own GPU
16 GB or more
Loads ComfyUI's template with room to work. 12 GB loads the smaller INT8 ConvRot file.
Computed 3 November 2026
A GPU by the hour
$0.17 an hour
An RTX 3090 holds it with room to work, on SaladCloud, or $0.50 on RunPod, which runs its own data centres. Any node; it bills while it idles.
Read 19 October 2026
On-demand prices before tax, read 19 October 2026.
What we measured
Z-Image Turbo, 4 images at 1024×1024, on rented cards in Finland on 20 October 2026. They peaked at 15.2 to 15.4 GB, against a largest stage of 12.7 GB: working memory comes on top.
Measured 20 October 2026 How we measure
About the model
Z-Image is Tongyi-MAI's open image model. Tongyi-MAI's 6-billion-parameter image model, one of the few here under Apache 2.0. Turbo (November 2025) generates in 8 steps; Base (January 2026) is the undistilled model for fine-tuning and longer sampling.
Turbo: 6.2B · Base: 6.2B Text encoder: Qwen3 4B Apache 2.0 Released November 2025
Good at
- Fast drafts: Turbo needs 8 steps where Base takes 28 to 50.
- Commercial use without conditions beyond the Apache notice.
- Images up to 2048 by 2048 pixels (Base).
Watch out for
- Tongyi-MAI's "fits comfortably within 16G" is about its own code. ComfyUI's template loads the BF16 files.
- The NVFP4 Turbo file is for Blackwell cards only.
Every Z-Image file ComfyUI loads
Sizes are the exact byte counts Hugging Face reports, read 3 November 2026. Files marked "template" are the ones ComfyUI's official workflow downloads.
| Part | Precision | Size | Version | File |
|---|---|---|---|---|
| Diffusion model | BF16 · template | 12.31 GB | Turbo | z_image_turbo_bf16.safetensors |
| Diffusion model | INT8 ConvRot | 6.20 GB | Turbo | z_image_turbo_int8_convrot.safetensors |
| Diffusion model | NVFP4 | 4.51 GB | Turbo | z_image_turbo_nvfp4.safetensors |
| Diffusion model | BF16 · template | 12.31 GB | Base | z_image_bf16.safetensors |
| Diffusion model | INT8 ConvRot | 6.20 GB | Base | z_image_int8_convrot.safetensors |
| Text encoder | BF16 · template | 8.04 GB | all | qwen_3_4b.safetensors |
| Text encoder | FP8 mixed | 5.63 GB | all | qwen_3_4b_fp8_mixed.safetensors |
| Text encoder | FP4 mixed | 3.48 GB | all | qwen_3_4b_fp4_mixed.safetensors |
| VAE | FP32 · template | 0.34 GB | all | ae.safetensors |
What Tongyi-MAI says it needs
- "fits comfortably within 16G VRAM consumer devices" (Z-Image-Turbo, the vendor's own inference code, source)
Those figures and this page measure different things. A vendor's figure covers its own script at its default resolution, working memory included. This page counts the weights each stage holds, which is exact, and leaves the working memory as the space that remains, because it depends on your resolution and attention kernel and nobody publishes it per model. If generation runs out of memory with a file that loads, lower the resolution before stepping down a precision. ComfyUI's README says it "can run even the biggest open source models on as low as 4GB vram + 8GB ram" by streaming weights (source), so "streams" means slower, not impossible.
LoRA training for Z-Image
No trainer we checked publishes a VRAM figure for training Z-Image yet (kohya's sd-scripts and musubi-tuner, ostris' ai-toolkit, diffusion-pipe, and Tongyi-MAI's own repo). When one does, it goes here.
Questions
- How much VRAM does Z-Image need?
- In ComfyUI's own workflow for Turbo, the largest stage is the diffusion stage at 12.7 GB, and the whole workflow downloads 20.7 GB. Working memory for generation comes on top and grows with resolution. Tongyi-MAI itself says: "fits comfortably within 16G VRAM consumer devices" (Z-Image-Turbo, the vendor's own inference code).
- Can I run Z-Image on 8 GB of VRAM?
- Yes, with the INT8 ConvRot file: its largest stage, the diffusion stage, is 6.5 GB, which loads on 8 GB with 1.5 GB left for working memory. That is little room to generate at full resolution.
- Can I run Z-Image on 12 GB of VRAM?
- Yes, with the INT8 ConvRot file: its largest stage, the text encoder, is 8.0 GB, which loads on 12 GB with 5.5 GB left for working memory.
- Can I run Z-Image on 16 GB of VRAM?
- Yes, with the BF16 file: its largest stage, the diffusion stage, is 12.7 GB, which loads on 16 GB with 3.3 GB left for working memory.
- Which Z-Image file should I download?
- The highest precision that loads fully on your card. On a 24 GB card that is the BF16 file (12.3 GB). ComfyUI's template downloads BF16 with the BF16 text encoder. FP4 files (NVFP4) are made for Blackwell cards (RTX 50, RTX PRO 6000, DGX Spark); on older cards, use FP8 or INT8.
- Files and sizes: Comfy-Org/z_image_turbo on Hugging Face, read from the Hugging Face API. Template files: ComfyUI's Z-Image guide.
- Model and licence: Tongyi-MAI/Z-Image-Turbo and the Apache 2.0.
- Card memory: manufacturers' specifications, as on each GPU page. The checker covers NVIDIA cards; ComfyUI also runs on AMD and Apple silicon, where file formats and speed differ and we have not sized them.
- Fit rule: a stage loads fully under 85% of usable memory and is tight up to 95%, the same margins the LLM pages use.
- Prices: each provider's own pricing page, on the dates beside them. Comfy Cloud's answer comes from ComfyUI's own template for that variant, which uses built-in nodes only, and shows only for the variants whose template we checked.