FLUX.1 [dev]
A solid pick for self-hosting and full data control.
Overview
FLUX.1 [dev] is a 12-billion-parameter rectified-flow transformer built by Germany's Black Forest Labs to turn text prompts into images. Output quality and prompt adherence rival closed-source alternatives, and the model runs locally on laptop-class hardware at Q4. The catch: its licence is non-commercial only, so production use requires a separate agreement with Black Forest Labs. Being open-weight and EU-based, it's a solid fit for local, on-premise experimentation without data leaving the organisation.
Skill profile
Not disclosed
Strengths
- Open-source and self-hostable
Limitations
- API pricing not disclosed
Who is it for
- you want to control cost or self-host
- you have GDPR constraints
Ideal use cases
- Internal creative prototyping
- Image-generation research
- Local workflow testing
- Non-commercial illustration
- Computer-vision R&D
Access & availability
Key specifications
Privacy
Run it locally
| Quantization | Disk | RAM / VRAM | Typical hardware |
|---|---|---|---|
| Q4 · recommended | 6.8 GB | 9 GB | 16GB PC / M1+ Mac / 8GB GPU |
| Q8 · balanced | 12.8 GB | 16 GB | 32GB PC / 24GB Mac / RTX 4070 Ti+ |
| FP16 · max quality | 24 GB | 28 GB | 64GB Mac / dual 24GB GPUs |
Estimates for a moderate context. Long contexts need more RAM (KV cache).
Deploy
Copy-ready commands generated from this card. Adjust context length and GPU count to your hardware.
OpenAI-compatible server for production on NVIDIA GPUs.
pip install vllm
vllm serve black-forest-labs/FLUX.1-dev \
--max-model-len 32768 \
--tensor-parallel-size 1 \
--dtype autoAdvanced data · for expertsArchitecture, modalities, detailed cost, full benchmarks▾
| Arena Elo | — |
| MMLU | — |
| GPQA | — |
| HumanEval | — |
| SWE-Bench | — |
| MATH | — |