What is FLUX.2 [dev]?
FLUX.2 [dev] was released by Black Forest Labs on 25 November 2025. It is a 32-billion-parameter rectified flow transformer paired with Mistral's 24-billion-parameter Mistral-3 vision-language model as its text encoder. One checkpoint does text-to-image, single-image editing and composition from up to 10 reference images, at up to 4 megapixels.
The weights are gated on Hugging Face under the FLUX [dev] Non-Commercial License v2.0, the same licence that covers FLUX.1 Kontext [dev]. Images you make may be used commercially, but running the model in any revenue-generating activity or anything that reaches end users needs a paid licence from BFL. It replaces FLUX.1 Kontext [dev] as BFL's large open editing model; the API-only FLUX.2 [pro], [flex] and [max] sit above it.
How much does FLUX.2 [dev] cost?
| Plan | Price | What you get |
|---|---|---|
| Weights | Free | FLUX [dev] Non-Commercial License v2.0 |
| fal | $0.012 per megapixel | Input and output megapixels both billed |
Weights free for non-commercial use; commercial self-hosting needs a BFL licence (price not published). Hosted on fal at $0.012 per megapixel of input and output; Artificial Analysis lists $24 per 1,000 images.
Is FLUX.2 [dev] worth it?
FLUX.2 [dev] is a benchmark more than a pick. Artificial Analysis fixes it at 1000 Elo as the reference point for its image-editing board, which puts it fiftieth of 79 models. BFL's own FLUX 3 Image Klein scores 1069 on the same board for $9 per 1,000 images, against $24 for FLUX.2 [dev], and Fibo 2.0 leads the open models at 1085.
What it still has is the ecosystem. The Hugging Face page lists 81 adapters, 41 fine-tunes and 18 quantised builds, and it drew 296,097 downloads in the last month, ahead of Qwen-Image-Edit-2511's 214,694. If your pipeline already runs on FLUX LoRAs, this is the checkpoint they target, and it handles up to 10 reference images at up to 4 megapixels, more than most open rivals.
The size is the problem. At 32 billion parameters the transformer alone is about 64GB in BF16, and the 24-billion-parameter text encoder adds roughly 48GB more. BFL's model card shows a 4-bit build on an RTX 4090 only by sending prompts to a hosted text encoder, and it publishes no VRAM figure.
The licence narrows it further. Outputs are yours to sell, but you may not use them to train a competing model, you must run content filters or review outputs before distribution, and using the model for paid work needs a commercial licence with no public price. BFL's own pricing page lists it as local only, with no hosted API, so production users go through fal or another host.
Where it earns its place
A two-person studio selling custom product mock-ups
Place a client's product into several reference scenes at print resolution
Call FLUX.2 [dev] on fal with the product shot and up to 10 references; a 1-megapixel edit from a 1-megapixel input costs about $0.024. Paid client work through the weights would need BFL's commercial licence, so the hosted route is the simpler one.
An ML researcher studying image-editing fine-tunes
Train and compare LoRAs on a widely used open base
Non-commercial research is allowed under the licence, and 81 published adapters give ready baselines. Plan for an 80GB GPU or more for BF16, or a 4-bit build with a remote text encoder on a 24GB card.
A game studio's art-tools engineer
Prototype an in-house concept-art editor before deciding what to license
Testing in a non-production environment counts as non-commercial, so the team can evaluate it free. Shipping it to artists means a paid BFL licence, at which point FLUX 3 Image Klein's terms and higher score deserve a direct comparison.
Before you commit
- Setup
- Accept the licence on Hugging Face, then load Flux2Pipeline in Diffusers or use BFL's reference repository; the sample uses 50 steps, with 28 noted as a good trade-off.
- Integrations
- Diffusers, ComfyUI, BFL's flux2 GitHub repository; hosted by fal, Replicate, Runware, Together AI, Cloudflare and DeepInfra.
- Data and privacy
- Self-hosted, images stay on your hardware. Licence: non-commercial use of the model, outputs usable commercially except to train competing models, content filters or human review required. The open inference code includes example pixel watermarking; signed C2PA metadata comes only from BFL's API.
- Cost at scale
- On fal, 100,000 one-megapixel edits with one-megapixel inputs cost about $2,400 a month. Self-hosting for commercial use needs a BFL licence whose price is not published.
- Limits
- No official VRAM figure; consumer GPUs need 4-bit weights and an offloaded or remote text encoder; Artificial Analysis ranks it 1000 Elo on editing.
What’s good
- Big LoRA and fine-tune ecosystem: 81 adapters, 41 fine-tunes
- Up to 10 reference images and 4-megapixel output
- Generation and editing in one checkpoint
- Diffusers, ComfyUI and several hosted providers
What’s not
- 1000 Elo, fiftieth on Artificial Analysis's editing board, 69 below FLUX 3 Image Klein
- About 112GB of BF16 weights with the text encoder
- Non-commercial licence; commercial self-hosting price not published
- Outputs may not train competing models; content filters required
Best for
Teams with an existing FLUX LoRA pipeline and data-centre GPUs who need multi-reference editing at up to 4 megapixels.
FLUX.2 [dev] alternatives
FLUX.2 [dev]: common questions
- Is FLUX.2 [dev] worth it?
- FLUX.2 [dev] is the image model everyone measures against, and on that same scale BFL's smaller, cheaper Klein now beats it. We rate it 3.5 out of 5.
- How much does FLUX.2 [dev] cost?
- Weights free for non-commercial use; commercial self-hosting needs a BFL licence (price not published). Hosted on fal at $0.012 per megapixel of input and output; Artificial Analysis lists $24 per 1,000 images.
- What are the downsides of FLUX.2 [dev]?
- 1000 Elo, fiftieth on Artificial Analysis's editing board, 69 below FLUX 3 Image Klein; about 112GB of BF16 weights with the text encoder; non-commercial licence; commercial self-hosting price not published; outputs may not train competing models; content filters required.
- Who is FLUX.2 [dev] best for?
- Teams with an existing FLUX LoRA pipeline and data-centre GPUs who need multi-reference editing at up to 4 megapixels.
- How much does FLUX.2 [dev] cost for a team or in production?
- On fal, 100,000 one-megapixel edits with one-megapixel inputs cost about $2,400 a month. Self-hosting for commercial use needs a BFL licence whose price is not published.
- Is FLUX.2 [dev] safe for company data?
- Self-hosted, images stay on your hardware. Licence: non-commercial use of the model, outputs usable commercially except to train competing models, content filters or human review required. The open inference code includes example pixel watermarking; signed C2PA metadata comes only from BFL's API.
- What are the best FLUX.2 [dev] alternatives?
- The closest alternatives we have reviewed are FLUX 3 Image Klein, Fibo 2.0, Qwen-Image-Edit-2511 and HunyuanImage 3.0 Instruct.
Answers checked Nov 7, 2026.