whatmodel.sh / image / FLUX.2 [dev]
FLUX.2 [dev]
↗ share
+ contribute
FLUX.2 [dev] is an open-weights image model from Black Forest Labs, released November 2025. Its published size is 32B parameters. Licence: flux-dev-nc.
Score within the LMArena text-to-image comparison set only.
$ how it compares
The image models ranked either side of FLUX.2 [dev], out of 197 with a score, then more from Black Forest Labs.
| # | model | score | where | ||
|---|---|---|---|---|---|
| 59 |  Qwen-Image-Edit-2509 | 59±3 | api $0.030/image +local | ||
| 60 |  HiDream-O1-Image | 58±2 | api $0.010/image +local | ||
| 61 | FLUX.2 [dev] | 58±1 | api $0.046/image +local | ||
| 62 |  Reve V1 | 58±4 | comparison only | ||
| 63 |  Z-Image (Base) | 58±2 | api $0.010/image +local | ||
| more from Black Forest Labs | |||||
| 32 | FLUX.2 [max] | 64±1 | api $0.070/image | ||
| 36 | FLUX.2 [pro] | 63±2 | api $0.030/image | ||
| 37 | FLUX.2 [flex] | 62±2 | api $0.050/image | ||
$ what it's good at
text to image▂▅▅▃▄▄█▃▁▂▁▁ #42 of 126 · 56% of its score · 9 tests
Beats Wan 2.7 Image, Wan 2.5 Text-to-Image Preview, Recraft V4 here, though they rank higher overall.
| test | result | vs the others in its run | counts |
|---|---|---|---|
| GenExam (relaxed overall, GPT-5 judge) | 42.3% ✎ | ▄█▂▂▂▃▆▄·▂·▂ beats 9/16 | 24% |
| LMArena text-to-image | 1145.8 ✎ | ▂▂▃▄▇█▄▂▄▂·▂ beats 41/75 | 15% |
| QC-T2I-Bench (English, HCQ total) | 89.5% ✎ | ▄····▄█▄▄█▄█ beats 10/10 | 4.8% |
| TIIF-Bench v3 (short prompts, TIIF-Evaluator) | 87.2% ✎ | ▂█····▄▄▄··▂ beats 8/11 | 4.4% |
| UltraText Bench (Composite, bilingual dense text rendering) | 41.0% ✎ | ▁▂▂█▃▁▂▃·▁·▁ beats 7/22 | 4.3% |
| FEPBench instruction faithfulness | 53.1% ✎ | ▄·█···▄▄▄▄▄▄ beats 4/8 | 2.0% |
| HPSv3 score of Cosmos 3 report text-to-image evaluation images | 11.3800 ✎ | ▄···█··▄·▄·▄ beats 2/5 | 1.0% |
| CVTG-102ch PNED (Chinese character rendering, Cosmos 3 report) | 68.7% ✎ | ▄·······▄█▄▄ beats 1/5 | 0.3% |
1 more test
| test | result | vs the others in its run | counts |
|---|---|---|---|
| CVTG-500L PNED (English long-prompt text rendering, Cosmos 3 report) | 85.0% ✎ | █······█████ beats 1/5 | 0.0% |
image to image / editing▁▁·▃▂▃█▃▃▄▁▁ #40 of 62 · 20% of its score · 3 tests
Beats GPT Image 1 Mini here, though they rank higher overall. Trails FLUX.2 [klein] 9B, Qwen-Image-Edit, LongCat-Image-Edit here, though they rank lower overall.
| test | result | vs the others in its run | counts |
|---|---|---|---|
| Edit-Compass (EN, overall) | 52.2% ✎ | ▃▃▂·▃▆▂█·▃·▃ beats 13/20 | 12% |
| Arena image edit | 1224.6 ✎ | ▂▁▁▂▆▇█▄▆▅▁▂ beats 14/53 | 5.9% |
| GEditBench v2 (overall Elo) | 1006.0 ✎ | ▃▂····▂█▂▃▂▃ beats 6/14 | 2.3% |
other tests6 not in a task · 23% of its score
| test | result | vs the others in its run | counts |
|---|---|---|---|
| VVRBench accuracy (10,000 verifiable synthetic-scene tasks) | 19.1% ✎ | ▃·█▅▃·▅▃···▃ beats 8/10 | 16% |
| Rapidata Benchmark.ai image benchmark (overall Elo) | 945.7 ✎ | ▁▁▁▄▃█▂▁▁▂▁▁ beats 9/34 | 2.4% |
| UniGenBench + Physical-AI prompts, All 1,170 (Cosmos 3 report) | 87.6% ✎ | ▄···█··▄···█ beats 3/5 | 2.1% |
| PaintBench precise editing (mIoU, 20-task average) | 4.6% ✎ | ▃·····█▅▃·▃▅ beats 4/9 | 1.3% |
| ViGoR-Bench zero-shot visual reasoning, image models (Result Avg) | 29.9% ✎ | ▂···▂·▂█▆▅·▂ beats 7/15 | 0.9% |
| Knowledge Visualization T2I (Brief Caption, English overall) | 36.9% ✎ | ▅·▃·▃··▃·█▃▃ beats 8/9 | 0.5% |
$ where to run it
Through an API
api · Cloudflare · $0.046/image · developers.cloudflare.com ↗ · price sourceCloudflare Workers AI; 4 output tiles × 28 steps × USD0.00041/output tile/step; zero input-image tiles. Workers plan fee excluded; published USD equivalent is rounded, service bills Neurons. Product: Workers AI text-to-image. API model: @cf/black-forest-labs/flux-2-dev. Settings: width=1024, height=1024, steps=28, input images=0, output tiles=4. Price basis: derived fixed-setting output cost using published rounded USD rates. Additional/excluded charges: Workers Paid plan base fee, input images and usage beyond this setting. Currency: USD.
Other ways to run it
version
- Run it as published · needs ~19.7 GB, NVIDIA only · from huggingface.co/blog
uv run --with diffusers --with transformers --with accelerate --with torch --with sentencepiece --with bitsandbytes python -c "import torch; from diffusers import Flux2Pipeline; p=Flux2Pipeline.from_pretrained('diffusers/FLUX.2-dev-bnb-4bit', torch_dtype=torch.bfloat16); p.enable_model_cpu_offload(); p('a cat holding a sign', num_inference_steps=28, guidance_scale=4).images[0].save('out.png')"
Official weights and documentation: huggingface.co/black-forest-labs ↗
$ how the score was worked out · GenExam: A Multidisciplinary… 24%, LMArena leaderboard dataset 21%, Verifiable Visual Rewards Tra… 16% and 11 more
Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works
$ improve this page
Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.
data updated · sources and licences · json