🆕 POCKET-Qwen3.8-Flash-Next — a 180B model running on a laptop with 8 GB VRAM + 32 GB RAM · 4.17 tok/s measured.

📚 Collections

▶ POCKET Models — this family (on-device, no GPU) Darwin Family · Aether Foundation · VKAE Accelerated

POCKET-Zimage-CPU

Pick your build →

Photoreal images in 46 seconds on a CPU only. No GPU. No CUDA. No Python.

🚀 Try it live, no install → — generating on a CPU-only box.

The POCKET lineup →

Why this exists

Every image model assumes you have a GPU. Most machines don't.

POCKET-Zimage-CPU is Z-Image-Turbo packaged so that a plain office PC — no graphics card, no CUDA, no Python environment — produces a photoreal 512×512 image in under a minute. One binary, three files, done.

Samples

_{3 steps · 48.6 s · the default we ship}

_{4 steps · 60.7 s · no visible gain}

_{Korean prompt · note it lost the count}

_{Prompt: a red apple on a wooden table, photorealistic — Korean: 나무 탁자 위에 놓인 빨간 사과, 사실적인 사진. Same seed, CPU only.}

Measured, not claimed

All numbers below are from our own runs. GPU count used: zero.

Resolution	Time	Sampling	VAE	Peak RAM
512 × 512	46.4 s	32.7 s	12.5 s	6.42 GB
512 × 512 (Korean prompt)	45.3 s	32.1 s	12.0 s	6.42 GB
1024 × 1024	192.7 s	135.1 s	55.8 s	6.76 GB

_{Intel Xeon Gold 6526Y ×2 (32 cores / 64 threads), 48 threads, Q4_0, 3 steps, --fa --vae-tiling. Single run per row.}

Korean prompts cost nothing extra — 45.3 s vs 46.4 s. Language is not a speed penalty here.

How it got 5.3× faster

We started at 244 seconds and ended at 46. Every step is measured:

Change	Time	Peak RAM
Default settings (20 steps)	244 s	8.16 GB
→ 4 steps	62.1 s	8.16 GB
→ `--fa` (flash attention)	59.5 s	8.18 GB
→ 3 steps	48.6 s	8.00 GB
→ `--vae-tiling`	46.4 s	6.42 GB

The big one is step count. Z-Image Turbo is distilled for few-step sampling, but the tool's default is 20. Using the default throws away 5× for nothing.

3 steps is the floor. At 4 and 3 we cannot tell the images apart. At 2 the surface collapses — water droplets and wood grain vanish and the texture turns cloth-like.

--vae-tiling is free. It cuts VAE time 16% and peak RAM by 1.27 GB. At 1024×1024 it is the difference between 13.3 GB and 6.76 GB.

Do not use every thread you have. On a 32-core / 64-thread box, 48 threads took 59.5 s and 64 threads took 108.8 s — 1.8× slower. Hyper-threads fight each other for the same physical cores.

Files

File	Size	What
`z_image_turbo-Q4_0-pocket.gguf`	3.51 GB	Diffusion model, 4-bit (VIDRAFT CPU build)
(bring your own) Qwen3-4B-Instruct-2507-Q4_K_M	2.58 GB	Text encoder — download
(bring your own) `ae.safetensors`	0.16 GB	VAE — download
Total	≈ 6.25 GB

The diffusion model here is VIDRAFT's own CPU build of Z-Image-Turbo: 3.51 GB, with no visible quality change.

Run it

Get a stable-diffusion.cpp binary (releases), then:

sd-cli \
  --diffusion-model z_image_turbo-Q4_0-pocket.gguf \
  --vae ae.safetensors \
  --llm Qwen3-4B-Instruct-2507-Q4_K_M.gguf \
  -p "a red apple on a wooden table, photorealistic" \
  --cfg-scale 1.0 --steps 3 --fa --vae-tiling \
  -t 8 -H 512 -W 512 -o out.png

Set -t to your physical core count — not your thread count.

Honest limits

It cannot render text. Any words inside the image come out garbled, in every language. If you need accurate text in an image, this is the wrong tool.
Korean prompts lose count. "A red apple" gives one apple in English and five or six in Korean. Korean has no articles, so the singular signal is weak for the encoder. Reproduced at both 20 and 3 steps, so it is the encoder, not the step count.
1024 × 1024 takes 3 minutes on the machine above. Slower CPUs scale accordingly.
Measured on a server CPU. Laptop and mini-PC numbers are not in yet.

Credits and licensing

Component	License	Author
Z-Image-Turbo (diffusion)	Apache-2.0	Tongyi-MAI / Hangzhou Tongyi Laboratory
Qwen3-4B-Instruct (text encoder)	Apache-2.0	Qwen, Alibaba
GGUF conversion (upstream)	Apache-2.0	leejet
stable-diffusion.cpp (runtime)	MIT	leejet

This repository redistributes a re-quantized copy of Z-Image-Turbo and keeps the original copyright notices. We did not train this model. What is ours is the CPU packaging, the 3-step setting, the re-quantization, and the measurements on this page.

	Runs on	Strength
POCKET-Image-Zimage	GPU (Python)	Faster, renders Korean text via glyph-init
POCKET-Zimage-CPU (this)	CPU only	No GPU, single binary

Different jobs. Use the first if you have a graphics card, this one if you don't.

🧩 The POCKET Family — On-device AI by VIDRAFT

Big models, small hardware. No GPU, no cloud.

Models

📦 POCKET-35B-GGUF — flagship, PC / server, no GPU
📦 POCKET-26B-GGUF — compact 26B
🇰🇷 POCKET-KR-GGUF — Korean, Android
🍎 POCKET-KR-MLX — Korean, iPhone / Mac
🌍 POCKET-EN-GGUF — English, phone / PC
💻 POCKET-Qwen3.8-Flash-Next-GGUF — 180B on a laptop (8 GB VRAM + 32 GB RAM)
🖼️ POCKET-Image-Zimage — character-perfect text in any image
🖥️ POCKET-Zimage-CPU — photoreal images on a CPU only

Demos & tools (Spaces)

🎨 POCKET-Image Studio — text-in-image, generate in-page
🖥️ POCKET-35B-CPU — 35B answering on a CPU
🖥️ POCKET-26B-CPU — 26B on a CPU
🖼️ POCKET-Zimage-CPU — image generation on a CPU

📚 Full POCKET collection

Downloads last month: -

GGUF

Model size

6B params

Architecture

Hardware compatibility

4-bit

Model tree for FINAL-Bench/POCKET-Zimage-CPU

Base model

Tongyi-MAI/Z-Image-Turbo

Quantized

(76)

this model

Space using FINAL-Bench/POCKET-Zimage-CPU 1

Collection including FINAL-Bench/POCKET-Zimage-CPU

POCKET-MODELs

Collection

12 items • Updated about 4 hours ago • 42

FINAL-Bench
/

POCKET-Zimage-CPU

🆕 POCKET-Qwen3.8-Flash-Next — a 180B model running on a laptop with 8 GB VRAM + 32 GB RAM · 4.17 tok/s measured.

📚 Collections