Short answer, before anything else: there is no FLUX 3 download. Black Forest Labs has not released FLUX 3 weights, has not announced a date for them, and the only official FLUX 3 access is a request-gated early access programme for the video capability.
If a site is offering you a "FLUX 3 model download," it is offering you something else — usually an older FLUX checkpoint with a new label, sometimes a repackaged file with an unclear licence, occasionally worse.
The useful part of this article is what comes next: the FLUX models you genuinely can download today, what each one is licensed for, and the hardware each needs.
Sourcing note: licences and hardware figures below come from Black Forest Labs' official GitHub repository, model pages, and licensing page. Nothing is taken from mirror sites or reuploads. Last verified July 28, 2026.
What this article solves
The pain point: "flux 3 download" is a query with a high scam surface. People searching it want local, offline, no-API-cost generation. The honest answer is "not this model, not yet" — and almost nobody publishing on this keyword says that, because it makes for a short page.
The differentiator: instead of stopping at "not available," this gives you the actual downloadable FLUX line-up with the licence terms and VRAM requirements side by side, so you can pick a model that fits your GPU and your commercial situation today.
Why there is no FLUX 3 download
BFL's launch plan for FLUX 3 has four stages. Open weights are the last one:
- Video and audio generation and editing via APIs and private weight access ("FLUX 3 Video")
- Action prediction through selected research and commercial partners ("FLUX-mimic and FLUX 3 Action")
- Image synthesis and editing via APIs and private weight access ("FLUX 3 Image")
- Open-weight access to a multimodal backbone ("FLUX 3 Dev")
Note the wording in stages 1 and 3: private weight access. That is a licensed, negotiated arrangement, not a download link. Stage 4 — the genuinely open one — has no announced date.
BFL's own history suggests patience. FLUX.2 [klein] was announced as "coming soon" open source in November 2025; by June 2026 klein models were shipping on-device with ASUS ProArt laptops. Open weights follow production APIs at this lab, not the other way round.
What you can download today
Everything below is published by Black Forest Labs on Hugging Face, with licences stated in their official GitHub repository.
The FLUX.2 family
| Model | Size | Licence | VRAM | Inference (RTX 5090) |
|---|---|---|---|---|
| FLUX.2 [klein] 4B | 4B | Apache 2.0 | 8.4 GB | ~1.2s |
| FLUX.2 [klein] 4B Base | 4B | Apache 2.0 | 9.2 GB | ~17s |
| FLUX.2 [klein] 9B | 9B | FLUX Non-Commercial | 19.6 GB | ~2s |
| FLUX.2 [klein] 9B Base | 9B | FLUX Non-Commercial | 21.7 GB | ~35s |
| FLUX.2 [dev] | 32B | FLUX non-commercial (commercial licence available) | High — see below | — |
Figures for the klein variants are BFL's own published numbers. "Base" models are undistilled foundation models — slower, but maximum flexibility for fine-tuning. The distilled variants are the ones built for sub-second generation.
FLUX.2 [dev] is the heavyweight: a 32B open-weight model that BFL describes as the most powerful open-weight image generation and editing model available, combining text-to-image with multi-image editing in a single checkpoint. For consumer GPUs, BFL points to an optimised fp8 reference implementation created with NVIDIA and ComfyUI.
The FLUX.1 family (still widely used)
| Model | Purpose | Licence |
|---|---|---|
| FLUX.1 [schnell] | Text to image | Apache 2.0 |
| FLUX.1 [dev] | Text to image | FLUX.1-dev Non-Commercial |
| FLUX.1 Fill [dev] | In/out-painting | FLUX.1-dev Non-Commercial |
| FLUX.1 Canny [dev] | Structural conditioning | FLUX.1-dev Non-Commercial |
| FLUX.1 Depth [dev] | Structural conditioning | FLUX.1-dev Non-Commercial |
| FLUX.1 Redux [dev] | Image variation | FLUX.1-dev Non-Commercial |
| FLUX.1 Kontext [dev] | Image editing | FLUX.1-dev Non-Commercial |
| FLUX.1 Krea [dev] | Text to image | FLUX.1-dev Non-Commercial |
The autoencoder weights are also released under Apache 2.0.
The licence question, in plain terms
This is where most local-AI projects quietly get themselves into trouble.
- Apache 2.0 models — FLUX.1 [schnell], FLUX.2 [klein] 4B and 4B Base, the VAE — are permissive. Commercial use is fine.
- Non-Commercial licence models — most of the [dev] line — are not for commercial use as downloaded. Personal projects, research, and evaluation are fine. Client work, a paid product, or ads for your business are not.
- Commercial use of the non-commercial models requires a licence from BFL. Their published tiers: Builder (FLUX.2 klein models, 10K images/month, 1 domain, 10 users, fine-tuning and LoRA rights), Platform (klein Base 9B plus FLUX.2 [dev], 100K images/month), Professional (FLUX.2 [dev], 100K images/month, up to 3 domains — aimed at agencies producing for named clients), and Enterprise (all models, custom terms). BFL also offers a synthetic-data licence with explicit rights to train on outputs for a contracted use case.
"I downloaded it from Hugging Face so it's free to use" is not how these licences work. The download is open; the commercial rights are not.
Not ready to build a local GPU pipeline? Flux 3 AI runs in the browser — prompt-to-image generation, reference workflows, editing, and upscaling with no install, no VRAM budget, and no licence homework. Try the image generator or upscale existing frames.
How to download safely
- Start from the official repo. black-forest-labs/flux on GitHub lists every open model with its Hugging Face link and licence. Use those links, not search results.
- Verify the Hugging Face org. Weights should come from
black-forest-labs. Community reuploads may be modified, may lack the licence file, and may not match the model card. - Read the model card's licence field before you generate anything commercial.
- Use the reference inference code. BFL's repo ships minimal inference code for image generation and editing, plus install instructions including a TensorRT path.
- Match the model to your GPU. Under ~10 GB VRAM, klein 4B is the realistic option. Around 20 GB, klein 9B opens up. FLUX.2 [dev] at 32B is a different tier of hardware — use the fp8 implementation if you are on consumer cards.
Warning signs on "FLUX 3 download" pages
- Any download link claiming to be FLUX 3 weights — they do not exist publicly
- A "FLUX 3 free download" that requires a survey, an installer, or a password-protected archive
- Claims of "FLUX 3 Pro" or "FLUX 3 Ultra" files — those tier names are FLUX.1/FLUX.2 vocabulary
- Repos with no licence file, or a licence copied from a different model
- Sites that describe FLUX 3 as an image generator only — the announcement is explicitly multimodal, video-first
FAQ
Can I download FLUX 3? No. Weights have not been released. FLUX 3 Dev is stage four of the launch plan with no date.
Is FLUX 3 free? No pricing or free tier has been published. Early access is by request.
What is the best free FLUX model I can run locally? For permissive commercial rights, FLUX.2 [klein] 4B (Apache 2.0, 8.4 GB VRAM) or FLUX.1 [schnell] (Apache 2.0). For maximum quality where the non-commercial licence is acceptable, FLUX.2 [dev].
Can I use FLUX.1 [dev] for client work? Not under the non-commercial licence as downloaded. You need a commercial licence from BFL.
How much VRAM do I need? BFL's published figures: 8.4 GB for klein 4B, 9.2 GB for 4B Base, 19.6 GB for klein 9B, 21.7 GB for 9B Base. FLUX.2 [dev] at 32B needs considerably more; use the fp8 reference implementation on consumer GPUs.
Will FLUX 3 weights be released eventually? BFL lists an open-weight multimodal backbone in the launch plan. No date has been given.
Bottom line
The download you searched for does not exist yet. The downloads that do exist are good — a 4B Apache 2.0 model that generates in about a second on a 5090, and a 32B open-weight model that BFL positions at the top of the open field.
Pick by licence first and VRAM second, and use the official repo as your only source. If local inference is not where you want to spend your time at all, generate in the browser instead and skip the whole stack.
Sources
- black-forest-labs/flux on GitHub — complete open-weight model list with licences, install and inference code
- Black Forest Labs on Hugging Face — official weight downloads
- FLUX.2 klein model page — VRAM, inference times, per-variant licences
- FLUX.2: Frontier Visual Intelligence (BFL, November 25, 2025) — FLUX.2 [dev] 32B, fp8 implementation with NVIDIA and ComfyUI
- FLUX open weights licensing — Builder / Platform / Professional / Enterprise and synthetic-data terms
- FLUX 3 — Real World Models (BFL, July 23, 2026) — launch plan, private weight access wording, no open-weight date
- FLUX 3 model page — current availability
- BFL API pricing — API alternative to local inference, and licensing tiers
- FLUX.2 klein on-device (BFL, June 4, 2026) — precedent for open-weight release timing
- BFL API documentation — hosted access to the same model family
Scope note: licence summaries here 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, and does not distribute model weights.
