Skip to content
← Models
Open weightsTongyi-MAI (Alibaba)/10 trustVerified
Z-Image-Turbo logo

Z-Image-Turbo

Alibaba's 6B Apache-2.0 image model — #1 open-source on Artificial Analysis, excellent in 8 steps with bilingual text rendering.

Visit official page →

About

Z-Image-Turbo (November 26, 2025) is Alibaba Tongyi's 6B-parameter distilled text-to-image model: DiT architecture with a Qwen3 text encoder, delivering high-quality images in 8 steps with sub-second latency on datacenter GPUs and fitting in 16GB VRAM for consumers. It ranked #1 open-source model (8th overall) on the Artificial Analysis text-to-image leaderboard (Dec 2025).

Apache 2.0 licensed — fully commercial-friendly, which is rare at this quality tier. Standout strengths: bilingual English/Chinese text rendering (especially Chinese characters, where most Western models fail) and strong instruction adherence. Three variants exist: Turbo (distilled, 8-step), Base (for fine-tuning), and Edit (instruction-based image editing). Runs on ComfyUI and diffusers; GGUF quants run on 6GB VRAM.

Specs

Context window

N/A (text-to-image model)

Pricing

Free — open weights (Apache 2.0), self-hosted, commercial use allowed.

Free quota

Free — download and run the weights yourself (Apache 2.0); GGUF quants for low VRAM.

API access

Self-hosted (diffusers, ComfyUI; Hugging Face and ModelScope checkpoints; GGUF quants).

Weights

Open weights

Released

November 26, 2025

Benchmarks

Independent: Artificial Analysis text-to-image leaderboard — 8th overall, #1 open-source (Dec 2025, crowdsourced human preference). Provider-claimed (technical report, arXiv 2511.22699): competitive quality at 8 NFEs — treat as provider-claimed.

Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.

Use cases

image-generationtext-to-imagedesignmarketinglocalization

Trust score — 8/10

/10 trustVerified

Scored 8/10 (verified tier). The rare combination of top-tier quality, Apache 2.0 commercial licensing, and efficiency (8 steps, 6B) — with independent leaderboard validation (#1 open-source on Artificial Analysis) and 629k HF downloads. Docked two points: (1) crowdsourced leaderboard positions are directional, not rigorous benchmarks; (2) the ecosystem (LoRAs, tutorials) is younger than FLUX/SDXL, and documentation is thinner.

How the trust score is built

Full methodology →
Adoption · stars/downloads, log-scaled vs peers
8.5/10
Maintenance · recency, releases, issue load
7.0/10
Openness · license + docs
9.1/10
Manual review · sanity read + red-flag check
8.0/10
  • · HF downloads 629,112 (peer-normalized)
  • · released 321 days ago
  • · manual review: rare quality+Apache-2.0+efficiency combo; #1 open-source on Artificial Analysis; ecosystem younger than FLUX/SDXL; license: Apache-2.0

Sources