The best open-weight image models in 2026: which one to run, and when

At a glance

• Top pick: Krea 2 Turbo. Best general-purpose realism you can run locally, and the deepest LoRA ecosystem. Free commercially under $1M company revenue.

• Best license: Z-Image Turbo. Plain Apache 2.0, 6B, most-downloaded open image model of the year. No negative prompts.

• "Open weight" stopped meaning "free to use." Read the license before the benchmark: revenue caps and non-commercial clauses are now the norm.

• The top-scoring open model is one most people cannot use. Ideogram 4.0 leads the open-weights arena, and it is non-commercial and gated.

What "open weight" actually means in 2026

This category got much busier this year and much less permissive at the same time. Krea, Ideogram, Baidu, Microsoft and NVIDIA all shipped open image weights between April and July 2026, and almost every one attached conditions. The conditions are where the real differences sit now.

So the ranking below is settled as much by license text and ecosystem depth as by output quality. A permissive model nobody trains LoRAs for won't do your job, and neither will the arena leader that needs three 80GB cards to load. Before you download, know that FLUX 3 is API-only and Alibaba's Qwen image line stopped publishing weights in December 2025, so Qwen-Image 2.0 and 3.0 are API-only whatever a few roundups claim. There is also no SD4. And NVIDIA has agreed to acquire Hugging Face, so keep local copies of what your work depends on.

How we picked

• Commercial use, in writing. Whether you can sell the output, and whether you can run the model for money. Different questions, and a model card's summary is not the license.

• What it takes to run. The real VRAM floor including quantized builds, and whether an Apple Silicon path exists.

• Ecosystem depth and liveness. LoRA counts, ControlNet availability, and last actual weight release.

• How we compared. Nothing was tested hands-on. This rests on license text, published specs and live counts, checked 10 September 2026.

Model comparison

1. Krea 2 Turbo - general-purpose realism, up to a revenue line

Krea's Krea 2 weights landed on 22 June 2026: a 12.8B transformer on a Qwen3-VL-4B text encoder. It became the default local model faster than anything before it, and Civitai now carries 6,384 Krea 2 assets, 5,946 of them LoRAs, in under three months. Most of the popular ones are realism and likeness work.

Two checkpoints are open, Raw and Turbo, both Medium-class, while Krea 2 Large stays API-only at more than twice the size. That costs less than you'd think: on the Artificial Analysis board today, Large ranks 23rd at 1024 Elo and the downloadable Medium Turbo 25th at 1023. The week-one problem is faces. It returns the same synthetic face repeatedly, realism LoRAs make that worse, and with no edit model and no true ControlNet a sketch-to-render workflow loses its control stack.

Best for: Photoreal work, product shots and character LoRAs, on 12GB or more.

License: Krea 2 Community v.1. Free below $1M company revenue, trailing twelve months, affiliates aggregated. You own the outputs. Above that, $1,000 per month per model as of September 2026, and no client work.

Runs on: 24.5GB at bf16, 12.2GB at fp8, 7.2GB at NVFP4, plus the text encoder. On a Mac, 24GB is the fp8 floor and bf16 wants nearer 48GB.

Watch out for: GGUF is broken in the standard ComfyUI plugin. Use a patched fork or the int8 convrot builds.

2. Z-Image Turbo - the fast one with nothing attached

Alibaba's Tongyi-MAI group released Z-Image Turbo on 26 November 2025 under plain Apache 2.0, and at 693,697 pulls in the last 30 days it's the most-downloaded open image model going. It's 6B at 8 steps, and the only model here whose own lab ships a ControlNet. Running CFG-free costs you negative prompts, silently dropped at guidance 0, so everything must be phrased positively.

Best for: Fast iteration and concepting on modest hardware.

License: Apache 2.0, unmodified. Nothing restricts commercial use of weights, derivatives or outputs.

Runs on: 16GB at bf16, down to 4GB with Q4_0 or Q3_K.

Watch out for: Turbo is not fine-tunable directly, skin and hair are weak, and Z-Image-Edit still hasn't shipped.

3. Qwen-Image-2512 - text inside images, and the least-bad editor

If your work puts words inside the picture, Qwen-Image-2512 is the one to run. This 20B transformer from 31 December 2025 holds dense poster layouts where the alternatives garble or mis-space them, Chinese typography genuinely works, and with Qwen-Image-Edit-2511 it's the least-bad open model at holding a face through an edit.

Best for: Posters, packaging comps, slides, UI mockups, instruction edits.

License: Apache 2.0, unmodified. Qwen Chat's 2025 non-commercial clause is gone from the May 2026 rewrite.  

Runs on: 58GB at bf16, so quantize. fp8 is 20.4GB and wants a 24GB card; Q4_K_M at 13.1GB is the realistic path.

Watch out for: The line stopped nine months ago and the fine-tune scene stalled at 1,163 Civitai LoRAs, only two of the 600 newest.

4. Anima - anime and illustration that fits on 8GB

CircleStone Labs built Anima with Comfy Org and shipped v1.0 in mid-May 2026. At 2B parameters it has taken over the illustration end of this field: 244 of the 600 newest Civitai checkpoints declare it as their base, 40.7% against 21.3% for Illustrious. Its Qwen3 0.6B text encoder is why, reading natural-language captions as well as Danbooru tags.

Best for: Anime and illustration described in plain sentences.

License: CircleStone Non-Commercial v1.2, plus NVIDIA's Open Model License via its Cosmos derivation. The model is non-commercial; outputs are explicitly yours to sell. Building it into a paid product is out.

Runs on: 6GB to 8GB per community testing. The checkpoint is 4.2GB.

Watch out for: Structural control is thin, with several adapter types still preview-era and no Canny at all.

5. FLUX.2 [klein] 4B - the license you don't have to read twice

Every other Black Forest Labs model worth running is non-commercial, which makes FLUX.2 [klein] 4B the useful one for a working creator. Released 15 January 2026 at 3.88B parameters, it's Apache 2.0 and the only klein weights you can download without accepting terms first. Text-to-image plus reference editing in one checkpoint, four steps, about 32 seconds for 1024px on an M1 Max through mflux. It also beats FLUX.1 [dev] on the open-weights board, 861 Elo to 843.

Best for: Reference-image editing, small-footprint local work, Apple Silicon.

License: Apache 2.0 and ungated, for the 4B builds only. klein 9B, the one most people run, is non-commercial.

Runs on: About 16GB of bf16 weights all in. A 4-bit MLX build is 4.6GB.

Watch out for: People. The family's faces need a second pass, and on a 3090 the fp8 builds save memory without speeding anything up. 

6. SDXL and Illustrious - where the control tooling still lives

Nobody should pick SDXL for output quality in 2026, and plenty of people should still have it installed. The ControlNet, IP-Adapter and inpainting stack built around it since 2023 is the deepest one, and people who otherwise moved to Krea 2 keep SDXL merges for high-resolution context-aware inpainting. Illustrious, its Danbooru-trained descendant, carries 196,000-plus LoRAs by Civitai's own count.

Best for: Structural control, pose locking, heavy inpainting, the oldest hardware here.

License: CreativeML Open RAIL++-M, the most permissive here: no revenue threshold, no registration, granted irrevocably. Illustrious v0.1 is Fair AI Public License 1.0-SD, copyleft and not non-commercial.

Runs on: 8GB, Stability's own stated figure, and lower with offloading.

Watch out for: Illustrious is genre-specific and its open weights stop at v2.0. NoobAI-XL's added terms cover "model-generated products."

Which one fits your desk

• Billing client work as a freelancer: Krea 2 Turbo until your revenue nears $1M, or Z-Image Turbo and klein 4B to never check a threshold again.

• On 8GB or an older card: Anima for illustration, Z-Image Turbo for everything else. Both are comfortable there.

• On a Mac: klein 4B through mflux, or Krea 2 Turbo in Draw Things at 32GB. Chip generation matters as much as memory.

• Doing sketch-to-render or heavy inpainting: keep SDXL or Illustrious for the control passes. Most modern models have not replaced ControlNet.

• Shipping a product or paid API: Z-Image Turbo, Qwen-Image-2512 or klein 4B. Anima forbids it outright.