FLUX AI: What It Is and Which Model to Use

"FLUX AI" gets used as though it names one product. It does not. It names a family built by Black Forest Labs that now spans three generations, five API tiers, several downloadable checkpoints, four task-specific tools, and a multimodal frontier model in gated early access.

That is why "is FLUX AI good?" has no answer and "which FLUX should I use?" has a very precise one.

Sourcing note: every model name, price, licence and capability below comes from Black Forest Labs' own model pages, API documentation, pricing page and announcements. Where BFL describes its own work in marketing terms, that is labelled as their claim. Last verified July 30, 2026.

What this article solves

The pain point: newcomers search a family name and get answers about whichever member the writer happened to use. Someone comparing FLUX to another tool ends up comparing against a random tier — often the cheapest or the oldest.

The differentiator: the whole family on one page with licence and price on every row, then a decision path that ends at exactly one model instead of "it depends."

Who makes it

Black Forest Labs — a research lab founded in 2024, based in Freiburg, Germany and San Francisco, roughly 70 people. Their about page describes the founding team as pioneers of Latent Diffusion, Stable Diffusion and FLUX.1. In December 2025 they raised a $300M Series B at a $3.25B post-money valuation.

Their strategy is "open core": genuinely open-weight models alongside paid API access and commercial weight licences.

The three generations

FLUX.1 (2024)

The generation that made the name. Text-to-image plus an editing suite — Fill, Canny, Depth, Redux, Kontext, Krea. FLUX.1 [schnell] is Apache 2.0; most of the rest is under a non-commercial licence. BFL now classes parts of this line as legacy.

FLUX.2 (November 25, 2025)

The current production family, and BFL's documented recommendation for all new work. What it added: up to 10 reference images in one request, editing at up to 4MP, markedly better text rendering, and stronger prompt following. Architecturally it couples a Mistral-3 24B vision-language model with a rectified flow transformer.

FLUX 3 (July 23, 2026)

The frontier model — multimodal, trained jointly on images, video and audio, plus action prediction. Generates video with native audio up to 20 seconds per generation. Currently request-gated early access, video capability first, with no published pricing.

The full line-up

Billing rules that change the maths: 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.

The tools people miss

Beyond generation, four task-specific endpoints — and three of them take no prompt at all:

• Erase — remove an object, rebuild what was behind it (needs a mask)
• Outpainting — extend an image past its borders (needs target dimensions)
• Deblur — sharpen a blurry photo (needs nothing)
• Virtual Try-On — put a garment on a person (needs a prompt)

---

Want to use FLUX-style workflows without picking a tier? Flux 3 AI is an independent browser workspace — generation, editing, reference workflows and upscaling with credits instead of API wiring. Open the image generator or see the plans.

---

Decision path

Answer in order and stop at the first yes.

1. Do you need video with sound? → FLUX 3, request early access. Nothing else in the family does it.
2. Must you self-host with commercial rights and no licence negotiation? → FLUX.2 [klein] 4B, or FLUX.1 [schnell]. Both Apache 2.0.
3. Do you need the strongest open weights and can license them? → FLUX.2 [dev], 32B.
4. Do you need the best hosted quality? → FLUX.2 [max].
5. Are you generating at volume via API? → FLUX.2 [pro], or klein 9B if throughput beats fidelity.
6. Is typography or step-count control the job? → FLUX.2 [flex].
7. Is your task erase / expand / deblur / try-on? → the matching Tool, not a generation model.
8. None of the above? → FLUX.2 [pro] is the sane default.

What FLUX is genuinely good at

Setting marketing aside, three things are documented rather than claimed:

• Text in images. BFL publishes specific techniques — quoted strings, placement, size, hex colour binding — and treats typography as a headline capability rather than a known weakness.
• Multi-reference consistency. Up to 10 references in one request is a structurally different way of working from single-image editing.
• Open weights that are actually usable. A 4B Apache 2.0 model needing 8.4 GB VRAM and running in about a second on a 5090 is a real deployment option, not a token gesture.

FAQ

Is FLUX AI free? Partly. FLUX.1 [schnell] and FLUX.2 [klein] 4B are Apache 2.0 and free to run and use commercially. Hosted API access is paid.

What is the best FLUX model? FLUX.2 [max] for hosted quality; FLUX.2 [dev] for open weights; FLUX 3 is the frontier but gated.

Is FLUX better than Midjourney or DALL·E? No neutral benchmark exists across these families, and pricing units differ enough that cost comparisons need your own workload to be meaningful.

Who owns FLUX? Black Forest Labs, an independent lab in Freiburg and San Francisco.

Can I use FLUX images commercially? Depends on the model and route. Apache 2.0 weights and paid API output are commercially usable; the non-commercial weight licences are not without a BFL licence.

What is the cheapest way to use FLUX? Self-host klein 4B (free, Apache 2.0), or call it via API at $0.014 for the first megapixel.

Is FLUX 3 available? Video capability in request-gated early access since July 23, 2026. Image early access was described as opening within weeks of that.

Bottom line

FLUX AI is a family with a clear internal logic: Apache 2.0 at the small end to seed the ecosystem, non-commercial weights in the middle to capture serious users, paid API at the top, and a multimodal frontier model gated while it stabilises.

Once you see that shape, choosing stops being hard. Decide on licence first, then price, and quality resolves itself — the tiers are ordered.

If you would rather skip the whole decision, use a browser workspace and come back to model selection when volume justifies it.

Sources

1. Black Forest Labs about page — team, locations, founding lineage
2. Laying the Foundations for Visual Intelligence — $300M Series B (BFL, December 1, 2025) — funding and valuation
3. FLUX.2: Frontier Visual Intelligence (BFL, November 25, 2025) — FLUX.2 capabilities and architecture
4. FLUX 3 — Real World Models (BFL, July 23, 2026) — multimodal frontier model and access status
5. BFL API pricing — per-megapixel rates and billing rules
6. FLUX.2 klein model page — Apache 2.0 variants, VRAM and latency
7. black-forest-labs/flux on GitHub — open model list and licences
8. FLUX open weights licensing — commercial licence tiers
9. BFL documentation overview — current recommendation and legacy classification
10. Official FLUX.2 prompting guide — documented text and colour techniques

Scope note: line-up and pricing as of July 30, 2026. Flux 3 AI is an independent creator workspace, not affiliated with Black Forest Labs.