What is Krea 2 (Turbo and Raw)?
Krea 2 Turbo and Krea 2 Raw were released as open weights on 22 June 2026, under the Krea 2 Community Licence dated the same day. Both are checkpoints of a diffusion transformer Krea trained from scratch; VentureBeat reports 12 billion parameters and a LocalAI Master setup guide 12.9 billion. Raw is an undistilled checkpoint taken partway through training of Krea 2 Medium, without aesthetic post-training, and runs at 52 steps with guidance 3.5 up to about 1K resolution. Turbo is distilled and post-trained, runs in 8 steps with guidance off and generates natively at up to 2K.
Krea's advice is to train LoRAs on Raw and run them on Turbo, and it published nine style LoRAs made that way. Comfy-Org's ComfyUI repack adds FP8, INT8, MXFP8 and NVFP4 builds alongside BF16, with a Qwen3-VL 4B text encoder and the Qwen-Image VAE. The larger Krea 2 Large and Krea's newer Krea 3 stay closed and run only in Krea's app and API.
How much does Krea 2 (Turbo and Raw) cost?
| Plan | Price | What you get |
|---|---|---|
| Weights | Free | Commercial use only while group revenue is under $1M a year |
| fal (Turbo) | $0.008 per megapixel |
Free weights under the Krea 2 Community Licence; commercial use only below $1M annual company revenue, an enterprise licence above. fal hosts Turbo at $0.008 per megapixel.
Is Krea 2 (Turbo and Raw) worth it?
Krea 2 is built for small teams that want to train their own look. The Raw checkpoint is an unpolished, undistilled base meant for fine-tuning, and LoRAs trained on it carry over to the fast Turbo checkpoint. Krea backs that split with an official repository, nine of its own style LoRAs and support in ComfyUI, Diffusers, the Ostris AI Toolkit and Kohya's musubi tuner.
On quality it sits in the upper middle. Artificial Analysis does not list the open checkpoints, but Krea's hosted Krea 2 Medium Turbo, which VentureBeat describes Turbo as a distilled version of, ranks 38th of 108 on the text-to-image board at 1020 Elo. That is above FLUX.2 [dev] at the board's fixed 1000 and below the open Qwen-Image-2.1 at 1045. Krea quotes about 2 seconds for a 2K Turbo image on consumer hardware it does not name.
Hardware needs are not published. LocalAI Master estimates about 13GB of weights in FP8, so 16GB of VRAM is a floor and 24GB comfortable, and about 26GB in BF16, which does not fit a 24GB card without offloading. Raw needs 52 guided steps against Turbo's 8, so it is a training base rather than a daily driver.
The licence is the real limit. Commercial use of the model, its derivatives or its outputs is free only while the whole company group earns under $1 million a year on a trailing twelve-month basis; past that you must stop and buy an enterprise licence. Every deployer must run content filters, redistributed derivatives must put 'Krea' at the start of their name, and the licence is revocable. Krea claims no rights in the images. Apache-licensed rivals such as Z-Image-Turbo and ERNIE-Image carry no revenue cap at all.
Where it earns its place
A three-person illustration studio with one RTX 4090
Train a LoRA on the studio's own style and produce client artwork at 2K
Train on Raw with the Ostris AI Toolkit or musubi tuner, then generate with Turbo in FP8 at 8 steps. Commercial use is free while the studio's revenue stays under $1 million, but it must still run a prompt or output filter.
A mobile app developer adding a stylised-image feature
Offer fast image generation without running GPUs
Call fal's Turbo endpoint at $0.008 per megapixel, so 100,000 images at about 1 megapixel cost roughly $800 a month, with no infrastructure to manage.
A machine-learning researcher studying fine-tuning
Experiment on a modern undistilled base without aesthetic tuning baked in
Raw ships without RLHF or aesthetic post-training, so the effect of a fine-tune is easier to isolate; run it at 52 steps with guidance 3.5 and plan for about 26GB of BF16 weights.
Before you commit
- Setup
- Clone krea-ai/krea-2 and run uv sync after accepting the gated download, or place Comfy-Org's repackaged files in ComfyUI; Turbo uses 8 steps with guidance 0.0.
- Integrations
- ComfyUI, Diffusers (Krea2Pipeline), SGLang and fal for inference; Diffusers, Ostris AI Toolkit, fal's trainer and Kohya's musubi tuner for LoRA training.
- Data and privacy
- Self-hosting keeps prompts and images local. The licence gives you ownership of outputs, makes content filters mandatory and requires a 'Krea' prefix plus the licence text on any redistributed derivative.
- Cost at scale
- On fal, 100,000 one-megapixel Turbo images cost about $800; self-hosting needs a 16GB-plus GPU in FP8 by LocalAI Master's estimate, and an enterprise licence once group revenue reaches $1 million.
- Limits
- Raw is trained to about 1K and is not meant for direct prompting; Turbo is tuned for 8 steps with guidance off; Krea publishes no VRAM figures or quality scores for the open checkpoints.
What’s good
- Turbo renders native 2K images in 8 steps
- LoRAs trained on Raw run on Turbo
- FP8, INT8 and NVFP4 builds for ComfyUI
- Hosted on fal at $0.008 per megapixel
What’s not
- Commercial use stops at $1M company revenue
- Content filtering is mandatory for every deployer
- No official VRAM figures or quality benchmarks
- Raw is trained only to about 1K resolution
Best for
Studios and independent developers under $1 million in revenue who want to fine-tune a house style and generate at 2K on their own GPUs.
Krea 2 (Turbo and Raw) alternatives
Krea 2 (Turbo and Raw): common questions
- Is Krea 2 (Turbo and Raw) worth it?
- Krea 2 Turbo makes 2K images in 8 steps on a gaming card and takes LoRAs trained on its Raw sibling, but a company past $1 million in revenue needs Krea's permission to sell the results. We rate it 3.8 out of 5.
- How much does Krea 2 (Turbo and Raw) cost?
- Free weights under the Krea 2 Community Licence; commercial use only below $1M annual company revenue, an enterprise licence above. fal hosts Turbo at $0.008 per megapixel.
- What are the downsides of Krea 2 (Turbo and Raw)?
- Commercial use stops at $1M company revenue; content filtering is mandatory for every deployer; no official VRAM figures or quality benchmarks; raw is trained only to about 1K resolution.
- Who is Krea 2 (Turbo and Raw) best for?
- Studios and independent developers under $1 million in revenue who want to fine-tune a house style and generate at 2K on their own GPUs.
- How much does Krea 2 (Turbo and Raw) cost for a team or in production?
- On fal, 100,000 one-megapixel Turbo images cost about $800; self-hosting needs a 16GB-plus GPU in FP8 by LocalAI Master's estimate, and an enterprise licence once group revenue reaches $1 million.
- Is Krea 2 (Turbo and Raw) safe for company data?
- Self-hosting keeps prompts and images local. The licence gives you ownership of outputs, makes content filters mandatory and requires a 'Krea' prefix plus the licence text on any redistributed derivative.
- What are the best Krea 2 (Turbo and Raw) alternatives?
- The closest alternatives we have reviewed are FLUX 3 Image Klein, Z-Image-Turbo, Qwen-Image-2.1 and ERNIE-Image (and Turbo).
Answers checked Nov 7, 2026.