What it is
Ideogram 4.0 launched on 3 June 2026 as the company's first model with downloadable weights. It is a 9.3-billion-parameter text-to-image model aimed at design: multilingual text in images, structured JSON prompts with bounding boxes and colour palettes, and native 2K output.
The weights, in fp8 and nf4 builds on Hugging Face, are licensed for non-commercial use only under the Ideogram 4 Non-Commercial licence. Only the inference code on GitHub is Apache 2.0, which is why some coverage calls the release open source. Commercial use runs through Ideogram's app and API, or a separate agreement.
Ideogram 4.5 followed on 30 September 2026 as an editing model built to avoid drift across repeated edits, available in the app and API with weights promised later.
Our review
For commercial teams Ideogram 4.0 is an API model with a research release attached. The weights on Hugging Face carry the Ideogram 4 Non-Commercial licence, and only the inference code on GitHub is Apache 2.0. The licence allows research, hobby use and testing in non-production environments; it rules out customer-facing deployment, generating images to promote revenue-earning products and fine-tuning for commercial use.
On quality, Artificial Analysis ranks 4.0 at Quality settings at 1026 Elo, 31st overall and third among open-weights models behind Qwen-Image-2.1 (1042) and Fibo 2.0 (1039). The default setting scores 1002. That board measures general preference, not lettering, which is where Ideogram competes. Its own pick of evidence is stronger: in a ContraLabs typography test designers chose it first 47.9% of the time, against 30.0% for Nano Banana 2, though Ideogram selected and published that comparison.
Self-hosting for evaluation has a catch. Plain prompts must be rewritten into the model's JSON format, and by default the reference code sends them to Ideogram's hosted API with your API key to do it. A team testing on confidential briefs is therefore sharing them with Ideogram unless it swaps in its own LLM using the published system prompt. The model also needs a separate Qwen3-VL-8B text encoder, and no VRAM figures are published for either build.
The hosted price is high for the ranking. Artificial Analysis lists $0.06 an image at default and $0.10 at Quality, about three times Nano Banana 2.1's $0.0336 for a model that sits 150 Elo points lower, and twenty times Z-Image Turbo's $0.005. fal's Instant variant costs $0.0075 but drops to 981. Ideogram's API agreement says it does not train on API inputs or outputs, which matters to agencies handling client copy.
Ideogram 4.5 High scores 1015 on the same board, below 4.0 at Quality, at the same $0.10. That says more about what the board rewards than about either model: neither is built to win a photorealism vote, and both are priced as if correct spelling were worth a premium.
Where it earns its place
A packaging design agency with eight designers
Produce label and poster mock-ups with exact copy
Through the API at about $0.10 per Quality image, 2,000 mock-ups a month cost about $200. JSON prompts pin text to bounding boxes, and the API agreement keeps client copy out of training.
A university researcher studying text rendering in diffusion models
Run and probe an open design model locally
The licence permits research use; the nf4 build needs a CUDA GPU and Diffusers. Swapping the magic-prompt step for a local LLM keeps the experiment offline and reproducible.
An ML engineer at an online print shop
Self-host image generation behind a customer-facing designer tool
The non-commercial licence rules this out without a separate agreement with Ideogram. The licensed path is the API, which defaults to 10 requests in flight, so peak traffic needs a raised limit.
Before you commit
- Setup
- Accept the Hugging Face gate and log in, pip install the GitHub repo, then run run_inference.py; an IDEOGRAM_API_KEY is needed for the default prompt rewriter. The API needs a prepaid balance at 1 cent a credit.
- Integrations
- Ideogram REST API; fal for Fast and Instant variants; Diffusers for the nf4 build; optional Hive moderation in the reference code.
- Data and privacy
- Weights: non-commercial only, no customer-facing use or commercial fine-tuning. API: Ideogram says it does not train on API inputs or outputs except content flagged for abuse. Self-hosted runs send prompts to Ideogram unless the prompt rewriter is replaced.
- Cost at scale
- 10,000 Quality images a month cost about $1,000 on Ideogram's API or $600 at default settings; fal's Instant variant would be about $75. Ideogram gives 14 days' notice of material price changes.
- Limits
- 256 to 2,048 pixels a side in multiples of 16, aspect ratios up to 6:1; the nf4 build is CUDA-only; 10 in-flight API requests by default; no published VRAM figures.
What's good
- Strong text rendering and JSON layout control
- Third among open-weights models on Artificial Analysis
- API agreement rules out training on API inputs and outputs
- Weights available for research and internal testing
What's not
- Weights licensed for non-commercial use only
- 31st overall on Artificial Analysis at Quality settings
- Reference code sends prompts to Ideogram's API by default
- $0.10 per Quality image, about three times Nano Banana 2.1
Best for
Designers and brand teams who need accurate text in images through an API, and researchers who want weights to study.
Alternatives
Ideogram 4.0: common questions
- Is Ideogram 4.0 worth it?
- Ideogram 4.0 is a typography specialist whose downloadable weights you can study but not ship, so for paid work it is an API model. We rate it 4.0 out of 5.
- How much does Ideogram 4.0 cost?
- API about $0.06 per image (default) or $0.10 (Quality), per Artificial Analysis; fal hosts Fast and Instant variants from $0.0075. App plans from $15/month yearly. Weights free for non-commercial use only.
- What are the downsides of Ideogram 4.0?
- Weights licensed for non-commercial use only; 31st overall on Artificial Analysis at Quality settings; reference code sends prompts to Ideogram's API by default; $0.10 per Quality image, about three times Nano Banana 2.1.
- Who is Ideogram 4.0 best for?
- Designers and brand teams who need accurate text in images through an API, and researchers who want weights to study.
- How much does Ideogram 4.0 cost for a team or in production?
- 10,000 Quality images a month cost about $1,000 on Ideogram's API or $600 at default settings; fal's Instant variant would be about $75. Ideogram gives 14 days' notice of material price changes.
- Is Ideogram 4.0 safe for company data?
- Weights: non-commercial only, no customer-facing use or commercial fine-tuning. API: Ideogram says it does not train on API inputs or outputs except content flagged for abuse. Self-hosted runs send prompts to Ideogram unless the prompt rewriter is replaced.
- What are the best Ideogram 4.0 alternatives?
- The closest alternatives we have reviewed are Qwen-Image-2.1, GPT Image 2.5 (Flare and Sunburst) and Midjourney V8.
Answers checked Oct 22, 2026.
Make Ideogram 4.0? Show this rating on your site
