Z-Image

Overview

Z-Image is Alibaba Tongyi-MAI's open-source image generation model family. The first checkpoint, Z-Image-Turbo, was released on November 26, 2025, per the project's own News log; the undistilled base model Z-Image followed on January 27, 2026.

The technical report opens by naming the problem it exists to solve, and it is worth quoting exactly:

"The landscape of high-performance image generation models is currently dominated by proprietary systems, such as Nano Banana Pro and Seedream 4.0. Leading open-source alternatives, including Qwen-Image, Hunyuan-Image-3.0 and FLUX.2, are characterized by massive parameter counts (20B to 80B), making them impractical for inference, and fine-tuning on consumer-grade hardware."

Z-Image's answer is an efficient 6B-parameter foundation generative model built upon a Scalable Single-Stream Diffusion Transformer (S3-DiT) architecture that challenges the "scale-at-all-costs" paradigm. The authors report completing "the full training workflow in just 314K H800 GPU hours (approx. $630K)" — an unusually specific and unusually low training-cost disclosure. The distilled Turbo variant offers "both sub-second inference latency on an enterprise-grade H800 GPU and compatibility with consumer-grade hardware (<16GB VRAM)."

That accessibility story is the whole model. Z-Image-Turbo has 933,580 Hugging Face downloads in the trailing 30 days (6,660,069 all-time) and 4,944 likes, making it the third most-downloaded text-to-image model on the Hub. It is also, verifiably, Apache 2.0 — not "open weights, restricted license." What it is not is the best image model available, and the Performance section below says so with numbers.

A note on naming

Three clarifications, because the family's own documentation invites confusion.

"Z-Image" names both the family and one specific checkpoint. Tongyi-MAI/Z-Image is the undistilled base model; Tongyi-MAI/Z-Image-Turbo is the 8-step distillation. They are different weights with different behaviour. arena.ai's modelKey uses the bare string z-image when it means Turbo — its row reads "modelKey":"z-image","modelDisplayName":"z-image-turbo".

Z-Image-Edit and Z-Image-Omni-Base do not exist as downloadable weights. Both are described in the model card, illustrated with showcase images, and Z-Image-Edit is named in the technical report's abstract ("our omni-pre-training paradigm also enables efficient training of Z-Image-Edit"). But the Model Zoo table has marked both "To be released" since launch, the GitHub repository has not been pushed to since February 9, 2026, and the Tongyi-MAI Hugging Face organisation holds exactly four repositories: Z-Image-Turbo, Z-Image, MAI-UI-8B, and MAI-UI-2B. Third-party sites that describe Z-Image-Edit as "released" are describing a model card, not a checkpoint.

Z-Image is not Qwen-Image. Both are Alibaba, both are Apache 2.0, both generate images — and they are separate model lines from separate teams. On arena.ai, Qwen-Image's qwen-image-2512 outranks Z-Image-Turbo.

Capabilities

• 8-step generation. Z-Image-Turbo "matches or exceeds leading competitors with only 8 NFEs (Number of Function Evaluations)." The quick-start sets num_inference_steps=9, which the model card notes "actually results in 8 DiT forwards," and guidance_scale=0.0 — Turbo runs without classifier-free guidance entirely.
• Bilingual text rendering. Alibaba claims Z-Image "excels at accurately rendering complex Chinese and English text" inside generated images, and the report calls bilingual text rendering one of two areas where results "rival top-tier commercial models."
• Photorealism. The second of those two areas. The blog phrases it as "Photography-level Realism."
• Sub-second latency. On an enterprise H800. The consumer-GPU path is slower but fits in under 16 GB of VRAM.
• Undistilled base for builders. The base Z-Image "preserves the complete training signal," "supports full Classifier-Free Guidance," "responds with high fidelity to negative prompting," and is positioned as "a good base for LoRA training, structural conditioning (ControlNet) and semantic conditioning."
• Prompt Enhancer. An optional reasoning pass that Alibaba says lets the model "transcend surface-level descriptions and tap into underlying world knowledge."
• Reward post-training. Turbo is the only variant with an RL stage, via the DMDR method (Distribution Matching Distillation meets Reinforcement Learning, arXiv 2511.13649).

Technical Specifications

• Hugging Face repositories: Tongyi-MAI/Z-Image-Turbo, Tongyi-MAI/Z-Image
• Architecture: Scalable Single-Stream Diffusion Transformer (S3-DiT). Text, visual semantic tokens, and image VAE tokens "are concatenated at the sequence level to serve as a unified input stream, maximizing parameter efficiency compared to dual-stream approaches."
• Parameters: 6B (vendor figure, referring to the DiT). The base repo's transformer shards total 12.31 GB in bf16 — consistent with ≈6.15B parameters.
• Transformer config: 30 layers, model dim 3840, 30 attention heads, 2 refiner layers, 16 input channels, QK-norm enabled
• Text encoder: a separate Qwen3ForCausalLM (hidden size 2560, 36 layers), not counted in the 6B
• VAE / scheduler: AutoencoderKL / FlowMatchEulerDiscreteScheduler
• Sampling — Turbo: 8 steps, no CFG (guidance_scale=0.0), no negative prompting, not fine-tunable
• Sampling — base: 28–50 steps, guidance_scale 3.0–5.0, negative prompting supported, fine-tunable
• Resolution: the base model card recommends 512×512 to 2048×2048 total pixel area, any aspect ratio. The Turbo quick-start generates at 1024×1024. Alibaba publishes no separate resolution range for Turbo, so this page does not state one.
• VRAM: <16 GB for Turbo (vendor claim)
• Download size: 20.55 GB (base, bf16 transformer); 32.90 GB (Turbo, fp32 transformer)
• Training cost: 314K H800 GPU hours, "approx. $630K"
• License: Apache 2.0, on both repositories and the GitHub LICENSE file

Alibaba's own Model Zoo table grades the variants rather than ranking them, and the trade-off is explicit: Turbo scores "Very High" visual quality but "Low" diversity and N/A fine-tunability; the base model scores "High" quality with "Medium" diversity and "Easy" fine-tunability.

Use Cases

• Local, offline generation on a consumer GPU — the model's reason for existing. No API key, no per-image cost, no content pipeline leaving your machine.
• Signage, posters, and UI mockups with embedded text — bilingual Chinese/English rendering is one of the two capabilities Alibaba claims parity with commercial models on.
• LoRA and ControlNet development — use the base checkpoint, not Turbo. Alibaba explicitly marks Turbo as non-fine-tunable.
• High-throughput batch generation — 8 forward passes per image with no CFG means roughly one-sixteenth the compute of a 50-step CFG sampler.
• Commercial products under a clean license — Apache 2.0 with no revenue cap, unlike Stable Diffusion 3.5's $1M Community License threshold.
• Distillation and few-step research — Decoupled-DMD (arXiv 2511.22677) and DMDR (arXiv 2511.13649) are published methods with open reference weights to study them against.

Performance / Benchmarks

Two independent arenas rate this model, and both are far less flattering than Alibaba's marketing. Note that arena.ai (the former LMArena) and Artificial Analysis are different organisations running different Elo scales; their numbers are not comparable to each other.

arena.ai text-to-image leaderboard (fetched 2026-07-08; vote cutoff July 5, 2026)

Z-Image-Turbo ranks #48 of 72 models, on 19,719 votes. Two facts follow from the table above, and both cut against the hype:

1. The entire top 11 is US labs — OpenAI, Meta, Reve, Google, Microsoft AI, xAI, and Ideogram. The best Chinese entry is Alibaba's proprietary qwen-image-2.0-pro at #12. On the closed image frontier, Chinese labs are not leading; on open weights, they are.
2. Z-Image-Turbo is not even Alibaba's best Apache-2.0 image model on this board. qwen-image-2512, also Apache 2.0, sits fifteen places higher.

Artificial Analysis text-to-image leaderboard (fetched 2026-07-08)

Artificial Analysis's own FAQ names Cosmos3-Super-Text2Image (agentic) as the current open-weights leader, at #8 overall.

* Artificial Analysis records no open-weights URL against its Z-Image Base row, which appears to be a gap in its data rather than a licensing fact — the checkpoint is public and Apache 2.0.

The vendor's own claims, and what happened to them

Alibaba's News log still carries, dated 2025-12-08: "Z-Image-Turbo ranked 8th overall on the Artificial Analysis Text-to-Image Leaderboard, making it the #1 open-source model!" That was true when written. It is no longer true. Seven months later the same leaderboard places Z-Image Turbo at #69.

Alibaba's other performance claim — "state-of-the-art results among open-source models" — is sourced to Alibaba AI Arena, an arena Alibaba itself operates. We attempted to verify it: aiarena.alibaba-inc.com renders entirely client-side and returns zero model names in its HTML. It is not first-party verifiable, and it is a vendor grading its own homework. We reproduce the claim and decline to endorse it.

What the numbers actually support: Z-Image is a strong, fast, permissively licensed model that fits on a consumer GPU. It is the most-downloaded text-to-image model of Chinese origin on Hugging Face. It is not the highest-rated image model, nor the highest-rated open one.

Limitations

• Popular ≠ best. #48 of 72 on arena.ai and #69 on Artificial Analysis. Downloads measure accessibility, not quality, and Z-Image's download lead is largely a function of its size and license.
• Turbo cannot be fine-tuned, guided, or negatively prompted. Alibaba's own table lists fine-tunability as N/A, CFG as ❌, negative prompting as ❌, and diversity as "Low" — distillation collapses seed-to-seed variety. Anyone building LoRAs must start from the base model, which is 44× less downloaded.
• Two of four advertised variants have never shipped. Z-Image-Omni-Base and Z-Image-Edit have been "To be released" since November 2025, and the repository has been untouched since February 9, 2026.
• The 16 GB VRAM figure is not the download size. The Turbo repository ships an fp32 transformer and totals 32.90 GB on disk.
• "6B" excludes the text encoder. The pipeline also loads a Qwen3 model with a 2560 hidden size and 36 layers. Real memory and disk footprints are larger than "6B" implies.
• All quality claims are vendor claims. The report's "performance comparable to or surpassing that of leading competitors" and the "state-of-the-art among open-source models" line rest on Alibaba's own arena, which cannot be independently read. Both independent arenas disagree.
• The Hugging Face card tags the model language: en despite the bilingual claim, and Alibaba publishes no text-rendering accuracy benchmark for either language.
• No first-party hosted API. Every live inference provider on Hugging Face is a third party, flagged isModelAuthor: false.

Pricing & Access

Self-hosting

Free, under a real open-source licence. Apache 2.0 on Tongyi-MAI/Z-Image-Turbo, on Tongyi-MAI/Z-Image, and in the GitHub LICENSE file. No revenue cap, no field-of-use restriction, no geographic exclusion. Commercial use, modification, and redistribution are permitted. Training data is not released, so this is open weights plus an open licence — not open data.

Hosted access

Hugging Face's inference-provider mapping lists three live hosts for Turbo. All three are third parties; Alibaba operates none of them, and each is marked isModelAuthor: false:

Replicate's PrunaAI deployment publishes per-image pricing tiered by resolution: $0.0025 up to 0.5 MP, $0.005 up to 1 MP, $0.01 up to 2 MP. This is a third-party price for a third-party deployment, not an Alibaba rate card.

Alibaba publishes no first-party hosted price for Z-Image that we were able to verify, so this page does not state one.

Free demos run on Hugging Face Spaces (Tongyi-MAI/Z-Image-Turbo, Tongyi-MAI/Z-Image) and on ModelScope.

Ecosystem & Tools

• diffusers — ZImagePipeline is merged into the main branch (PRs #12703 and #12715). The canonical inference path.
• ComfyUI — native support; ZImage classes appear in comfy/supported_models.py. A third-party ComfyUI-ZImageLatent node supplies the official resolutions.
• stable-diffusion.cpp — a pure C++ engine that the project says runs Z-Image "on machines with as little as 4GB of VRAM," across CUDA and Vulkan.
• DiffSynth-Studio — LoRA training, full training, distillation training, and low-VRAM inference.
• vllm-omni and SGLang-Diffusion — production serving paths.
• Cache-DiT — nearly 4× speedup on 4 GPUs "with negligible precision loss"; also covers Z-Image-ControlNet.
• Candle — Hugging Face's Rust ML framework supports Z-Image.
• MeanCache — training-free acceleration, up to 3.7× (China Unicom).

Community & Resources

• Z-Image: An Efficient Image Generation Foundation Model with Single-Stream Diffusion Transformer — the technical report (v1 2025-11-27, v5 2026-07-06); source of the 6B, 314K GPU-hour, $630K, and <16GB figures
• Tongyi-MAI/Z-Image on GitHub — 11,697 stars, Apache 2.0, the News log with both release dates
• Tongyi-MAI/Z-Image-Turbo on Hugging Face — Turbo weights, Model Zoo table
• Tongyi-MAI/Z-Image on Hugging Face — undistilled base weights
• Z-Image project blog — the official landing page
• Decoupled DMD: CFG Augmentation as the Spear, Distribution Matching as the Shield — the 8-step distillation method
• Distribution Matching Distillation Meets Reinforcement Learning — DMDR, Turbo's RL post-training
• arena.ai text-to-image leaderboard — where Z-Image-Turbo ranks #48
• Artificial Analysis text-to-image leaderboard — where it ranks #69
• Compare with Stable Diffusion 3.5, Qwen-Image 2.0, HunyuanImage 3.0, HiDream-O1-Image, and Seedream 5.0