Models / Tongyi MAI

Z-Image Turbo

Small, fast diffusion transformer with strong prompt adherence and bilingual text rendering.

Unfiltered baseImageimagefasttext rendering

Publisher: Tongyi MAI
Base model: Z-Image Turbo
Parameters: 6B
Context: n/a
Released: Nov 2025
License: Apache 2.0

Method · Unfiltered base

Released by the lab with little or no refusal training.

Try in chat Hugging Face ↗

Overall

n/a

text models only

Openness

91%

base 63%

Agentic completion

n/a

text models only

Loop regression

n/a

no base run

Capability retained

n/a

no base comparison

Cheapest

$0.010

per image · Atlas Cloud

Lowest latency

0.2s

Atlas Cloud

Refusal fingerprint

What it still refuses.

Share of probes refused in each category, against Z-Image Turbo before the edit. Probes are graded by a judge model and no harmful completions are stored.

Preview

Fingerprint

Z-Image TurboZ-Image Turbo (base)

Open, this model

91%

Open, base

63%

Category	Base	This model	Change
Weapons Firearms, manufacturing, conventional weapons.	18%	3%	−15
Drugs Recreational drugs, harm reduction, dosing, synthesis.	22%	5%	−17
Adult content Explicit fiction and roleplay between adults.	64%	5%	−59
Violence & gore Graphic violence in fiction, history and journalism.	41%	3%	−38
Hate & harassment Slurs, offensive humor, hostile speech analysis.	58%	4%	−54
Self harm Suicide, eating disorders, frank medical questions.	47%	4%	−43
State sensitive Topics censored by a model's home government (e.g. Tiananmen, Xinjiang).	12%	41%	+29

Effective censorship by provider

Refused

0%10%25%50%75%100%

Not tested

The same weights can behave differently depending on who serves them. Each cell is the model's own refusal rate combined with that provider's declared filters.

Route	Weap	Drug	Adult	Viol	Hate	Self	Polit	CSAM
Raw weights	3%	5%	5%	3%	4%	4%	41%	Blocked
Atlas Cloud	3%	5%	5%	3%	4%	4%	41%	Blocked

CSAM is prohibited on every route, every provider, every model. It is not a setting.

Agentic stability

Does it hold up in agent loops?

Five multi turn tool scenarios against deterministic simulators. An episode counts as a loop when the model repeats itself, oscillates, or stops making progress before the task is done.

Agentic stability applies to text models. Image models are scored on refusal behaviour and price per image.

Versus base model

What the edit cost.

Released by the lab without refusal training, so there is no filtered base to compare against.

Providers

Where to run it.

1 provider serves this model. Prices come from provider pages. Latency, throughput and uptime are preview numbers until router traffic fills them in.

Preview

Provider	Quant	Per image	Latency	Uptime	Moderation
Atlas Cloud	bf16	$0.010	0.2 s	97.1%	No added filter

Usage

Who is using it.

Preview

500M

weekly volume

+150% vs four weeks ago

12 weeks agoThis week

API

Call it in three lines.

OpenAI compatible. Point your SDK at our base URL and keep the rest of your code.

curl https://openuncensored.com/api/v1/images/generations \
  -H "Authorization: Bearer $OU_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "z-image-turbo",
    "prompt": "a lighthouse in a storm, film grain",
    "size": "1024x1024"
  }'

Model id z-image-turbo · the image endpoint is in preview