FLUX Schnell vs Dev vs Pro: Which to Use?

The FLUX.1 tier names have outlived the generation they were built for. People still search flux schnell and flux dev far more than they search the FLUX.2 variants that replaced them — partly because the local-AI ecosystem standardised on those checkpoints, and partly because the naming is genuinely sticky.

So this article does two jobs: settle the schnell / dev / pro decision on the facts, and show you what each tier corresponds to in the family Black Forest Labs actually recommends now.

Sourcing note: parameter counts, distillation methods, and licences come from BFL's official Hugging Face model cards and GitHub repository. Prices and VRAM figures come from BFL's pricing and model pages. Where a claim would need a benchmark nobody has published, this article gives the decision rule instead of a fake number. Last verified July 28, 2026.

What this article solves

The pain point: most schnell-vs-dev comparisons are about speed, when the decision is almost always about licence. People pick dev for quality, build a product on it, and discover the non-commercial licence after the fact.

The differentiator: licence is the first column of every table here, not a footnote — and the article maps the old tiers onto the current FLUX.2 line-up so you are not optimising within a generation you should have left.

The three tiers at a glance

Both open models are 12 billion parameter rectified flow transformers, per their official model cards. They are the same size — the difference is what was distilled into them and what you are allowed to do with the output.

The decision, in one question

Are you making money from the images?

• Yes, and you are self-hosting → FLUX.1 [schnell] (Apache 2.0) or, better, FLUX.2 [klein] 4B (also Apache 2.0, newer). Free and commercially usable, no licence conversation.
• Yes, and you want top quality → API access (FLUX.2 [pro] or [max]), or a BFL commercial licence for a dev-class model.
• No — personal, research, evaluation → FLUX.1 [dev] or FLUX.2 [dev]. Best open quality, non-commercial terms.

Everything else — speed, VRAM, step counts — is optimisation inside one of those three answers. Getting this order backwards is the single most common expensive mistake with FLUX models.

Where each tier genuinely wins

FLUX.1 [schnell]

The permissive one. Apache 2.0 covers the weights and BFL also releases the autoencoder under Apache 2.0, which matters if you are shipping a pipeline rather than just generating images.

Trained with latent adversarial diffusion distillation, it is built for speed. BFL's own model card claims cutting-edge output quality and competitive prompt following against closed-source alternatives — vendor framing, but the licence is objective and it is the reason schnell shows up in so many products.

Choose it when: commercial rights and cost-per-image dominate, and "good" beats "best."

FLUX.1 [dev]

The quality-per-download leader of its generation. BFL positions it as second only to FLUX.1 [pro] in output quality. It is also, by a wide margin, the checkpoint the fine-tuning and LoRA community built on — which is a real practical advantage that no spec sheet shows.

The catch is the licence. The FLUX.1-dev Non-Commercial License means client work, paid products, and business marketing are out unless you license it from BFL. Downloading it from Hugging Face grants access, not commercial rights.

Choose it when: you are doing personal or research work, or you hold a BFL commercial licence, and the existing LoRA ecosystem matters to you.

FLUX.1 [pro]

No weights, API only, commercial use included. You trade control for quality and zero infrastructure.

Choose it when: you want production output without a GPU pipeline — though for new projects, BFL's documentation now points you at FLUX.2 rather than the FLUX.1 pro line, and classifies FLUX1.1 [pro] models as legacy.

---

Neither hosting nor licence-reading appeals? Flux 3 AI runs the whole thing in a browser — generation, editing, reference workflows, upscaling — with credits instead of GPUs. Try the image generator or see what a credit pack covers.

---

The upgrade path: what these tiers are now

If you are starting fresh in 2026, the equivalent choices live in FLUX.2:

The FLUX.2 generation also adds things FLUX.1 simply cannot do: up to 10 reference images for character, product and style consistency, editing at up to 4MP, and markedly better text rendering. If your reason for staying on FLUX.1 is inertia rather than a LoRA library, the move is worth it.

Cost comparison for hosted use

Published BFL API pricing, per first megapixel:

• FLUX.2 [klein] 4B — $0.014
• FLUX.2 [klein] 9B — $0.015
• FLUX.2 [pro] — $0.03
• FLUX.2 [flex] — $0.05
• FLUX.2 [max] — $0.07

Billing notes that matter at volume: 1 MP = 1024×1024, resolution rounds up to the next MP separately for the output and each reference image, and every reference image counts as at least 1 MP. A multi-reference edit is therefore several times the price of a plain generation — budget accordingly.

FAQ

Is FLUX schnell free for commercial use? Yes. FLUX.1 [schnell] is Apache 2.0, as is the autoencoder.

Is FLUX dev free for commercial use? No. It carries the FLUX.1-dev Non-Commercial License. Commercial use requires a licence from BFL.

What is the difference between schnell and dev? Both are 12B rectified flow transformers. Schnell is distilled for speed with a permissive licence; dev targets higher quality under a non-commercial licence.

Can I download FLUX pro? No. FLUX.1 [pro] is API-only, no released weights.

Which is best for LoRA training? FLUX.1 [dev] has the largest existing ecosystem. For new work under permissive terms, FLUX.2 [klein] Base variants are designed for fine-tuning, and BFL licence tiers include fine-tuning and LoRA rights.

Should I still use FLUX.1 at all? Only for ecosystem reasons. BFL's documentation recommends FLUX.2 for all new projects.

What about FLUX 3? Different question entirely — FLUX 3 is a multimodal video-first model in gated early access, with no public weights or pricing.

Bottom line

Pick on licence first: schnell if you are selling the output and self-hosting, dev if you are not selling it (or you hold a licence), pro/API if you want no infrastructure.

Then upgrade the tier: klein 4B replaces schnell with better economics, FLUX.2 [dev] replaces dev with 32B quality and multi-reference editing, and FLUX.2 [pro] or [max] replaces the pro line.

If none of that sounds like a good use of your week, generate in the browser instead and let the model selection be somebody else's problem.

Sources

1. FLUX.1-schnell model card (Hugging Face, Black Forest Labs) — 12B rectified flow transformer, distillation method, Apache 2.0 licence
2. FLUX.1-dev model card (Hugging Face, Black Forest Labs) — 12B parameters, quality positioning, non-commercial licence
3. black-forest-labs/flux on GitHub — complete open model list with licences, autoencoder terms
4. BFL API pricing — per-megapixel pricing and billing rules
5. FLUX.2 klein model page — variant licences, VRAM, inference times
6. FLUX.2: Frontier Visual Intelligence (BFL, November 25, 2025) — FLUX.2 [dev] 32B, multi-reference, 4MP editing
7. FLUX open weights licensing — commercial tiers, fine-tuning and LoRA rights
8. BFL API documentation — current recommendation, legacy model classification
9. FLUX.2 pro / flex model page — production positioning and reference control
10. FLUX 3 — Real World Models (BFL, July 23, 2026) — status of the newest generation

Scope note: licence summaries are informational, not legal advice — read the licence text on the model card before commercial use. Flux 3 AI is an independent creator workspace, not affiliated with Black Forest Labs.