FLUX.1 Dev: specs, benchmarks, and how to run it on Lyceum

Get started: call FLUX.1 Dev on Lyceum

You can generate images with FLUX.1 Dev using Lyceum Technology's Serverless Inference API. Because this is a dedicated image generation model, you must use the standard images/generations endpoint via a direct HTTP POST request with Bearer authentication. Do not use the OpenAI chat completions endpoint.

import requests
 
 # BASE_URL is the base URL shown in your Lyceum dashboard.
 response = requests.post(
 f"{BASE_URL}/images/generations",
 headers={"Authorization": "Bearer <your lyceum api key>"},
 json={"model": "lyc-flux-1-dev", "prompt": "a sunset over the ocean", "aspect_ratio": "1:1"},
 )
 print(response.json()["image_url"])

Pricing and region for FLUX.1 Dev

FLUX.1 Dev runs in the Lyceum eu-north1 region, keeping processing in European data centres in Spain, Paris and the Nordics. The model is priced at a flat rate of $0.005 per image. Because the API operates on a serverless, pay-per-use model, there are no base fees, no minimum commitments, and no idle compute costs. You are billed exclusively for the successful image generations you request.

By leveraging Lyceum's infrastructure, your engineering team can bypass the complexities of provisioning high-VRAM instances, configuring CUDA environments, or managing container orchestration. The API handles the heavy lifting of model loading and execution, returning the generated image URL directly in the response payload.

What FLUX.1 Dev is good at

Strong photorealism and anatomy

FLUX.1 Dev excels at generating highly realistic images that overcome the typical "AI look" associated with older diffusion models. It is particularly strong at rendering complex human anatomy, including hands, faces, and dynamic poses, which have traditionally been failure points for text-to-image models. The 12 billion parameter rectified flow transformer architecture provides the massive capacity needed to capture intricate visual details, lighting nuances, and natural textures.

Exceptional prompt adherence

Unlike earlier generation models that often ignore complex instructions or blend concepts together, FLUX.1 Dev closely follows detailed, multi-subject prompts. It accurately maps spatial relationships, specific colors, and distinct attributes to the correct subjects within the generated image. This makes it highly effective for professional workflows that require precise visual outputs based on strict creative briefs, such as advertising mockups or concept art.

In-image text generation

One of the standout capabilities of the FLUX.1 family is its ability to generate legible, accurate text directly within images. Whether you need a storefront sign, a product label, or stylized typography integrated into a graphic design, FLUX.1 Dev handles text rendering with a high degree of reliability. This significantly reduces the need for post-generation editing and allows teams to automate the creation of marketing assets that include specific copy.

Limitations and trade-offs

High VRAM requirements for self-hosting

At 12 billion parameters, FLUX.1 Dev is a massive model. Running it at full precision (FP16) requires over 24GB of VRAM, making it difficult to deploy on standard consumer hardware without aggressive quantization. While it can be quantized to FP8 to fit on smaller GPUs, doing so can introduce slight quality degradations in fine details. For production workloads, self-hosting typically requires high-end data center GPUs like the NVIDIA H100 or A100, which can be expensive to provision and maintain. A newer model in the same family is covered in our notes on FLUX.2 Klein and how to run it.

Inference speed and latency

Because FLUX.1 Dev requires 30 to 50 inference steps to produce optimal, high-fidelity results, it is inherently slower than its distilled sibling, FLUX.1 Schnell (which requires only 1 to 4 steps). Generating a high-resolution image can take several seconds depending on the underlying hardware and optimization stack. Teams building real-time or highly interactive applications may need to implement aggressive caching, asynchronous background processing, or opt for the faster Schnell variant if latency is the primary constraint.

Non-commercial license restrictions

FLUX.1 Dev is released under the FLUX.1 [dev] Non-Commercial License from Black Forest Labs. The license claims no ownership of outputs and says you may use output for any purpose, including commercial purposes, with one carve-out: output may not be used to train, fine-tune or distill a model competitive with FLUX.1 Dev. The weights are the restricted part. Revenue-generating use of them, or use that has an impact on end users, falls outside the non-commercial grant and needs a separate license requested from Black Forest Labs. For fully open commercial use cases that involve hosting the weights, developers often look to the Apache 2.0 licensed FLUX.1 Schnell.

Benchmarks and how it compares

FLUX.1 Dev benchmark results

FLUX.1 Dev was among the strongest open-weights image models when Black Forest Labs released it on 1 August 2024, and it still scores above FLUX.1 Schnell and Stable Diffusion 3.5 Large. A newer generation has since passed it. On the Artificial Analysis Text-to-Image Arena, which uses blind human preference voting to calculate Elo scores, FLUX.1 Dev sits at 1028 on a board of 147 models led by GPT Image 2 (high) at 1339, with FLUX.2 Dev the highest-scoring open-weights entry at 1154.

Source: Artificial Analysis Text-to-Image Arena leaderboard, read 3 August 2026. Elo ratings move as models are added to the board.

Comparison to sibling models

Compared to FLUX.1 Schnell, the Dev variant produces noticeably higher quality images with better prompt adherence and finer details, but requires significantly more compute time (30-50 steps vs. 1-4 steps). Compared to the closed-source FLUX.1 Pro, Dev offers near-identical quality, as it is guidance-distilled directly from the Pro version - but provides the transparency of open weights for researchers and developers who want to inspect the architecture or build custom workflows.

Using it in production

Production configuration for FLUX.1 Dev

When deploying FLUX.1 Dev in production, managing latency and infrastructure costs are the primary challenges. Because the model requires 30 to 50 steps for high-quality generation, it is highly compute-intensive. Lyceum Technology abstracts this complexity by serving the model via our Serverless Inference API, allowing you to bypass the need to provision, optimize, and manage dedicated H100 or A100 instances yourself.

API parameters and aspect ratios

The Lyceum images/generations endpoint, called at the base URL shown in your Lyceum dashboard, accepts standard parameters such as prompt and aspect_ratio. For FLUX.1 Dev, it is highly recommended to keep the prompt descriptive and specific, as the model's strong prompt adherence will attempt to render every detail you specify. You can adjust the aspect ratio (e.g., 1:1, 16:9, 9:16) to suit your application's UI requirements, whether you are generating square profile pictures or widescreen marketing assets.

Pricing economics and scaling

On Lyceum, FLUX.1 Dev is priced at a flat rate of $0.005 per image. FLUX.1 Dev is priced per image, focusing purely on high-quality generation. For a production application generating 10,000 images per month, the total inference cost would be exactly $50.00. Because the API is serverless, you do not pay for idle GPU time between requests. All generation occurs in the eu-north1 region, ensuring predictable performance without the overhead of maintaining a persistent GPU cluster.

Running FLUX.1 Dev on EU-sovereign infrastructure

Why run FLUX.1 Dev on Lyceum

For European AI startups and enterprise teams, data privacy and regulatory compliance are non-negotiable. Many popular image generation APIs route requests through US-based servers, creating potential GDPR and data residency risks. Lyceum solves this by serving FLUX.1 Dev from the eu-north1 region, in European data centres in Spain, Paris and the Nordics.

GDPR compliance and data residency

When you generate images using Lyceum's Serverless Inference API, your prompts and generated outputs never leave Europe. That means GDPR-compliant processing in European data centres for sensitive creative briefs, proprietary product concepts, or user-generated content. We never train on customer data, and inference prompts and outputs are not retained after processing. Lyceum states no EU AI Act conformity position, so AI Act obligations remain yours to assess for your own application.

OpenAI-compatible integration and cost advantages

Switching to Lyceum requires zero architectural changes if you are already using standard API patterns. By updating your endpoint URL and providing a Lyceum API key, you can immediately start generating images with FLUX.1 Dev. Furthermore, Lyceum serves the model from European data centres in Spain, Paris and the Nordics, with GPU compute billed per second and no base fee. If your team eventually needs to fine-tune models or run custom workloads, you can provision a dedicated GPU VM that bills per second of use.

To learn more about our infrastructure approach, read our guide on GDPR-compliant AI inference in Europe.

Sources

[1] FLUX.1 Dev Model Card - Hugging Face; [2] FLUX.1 Dev Model by Black Forest Labs - NVIDIA NIM; [3] Announcing Black Forest Labs and FLUX.1, 1 August 2024; [4] FLUX.1 GitHub Repository