You searched "FLUX 3 ComfyUI," and the honest answer disappoints before the tutorial starts: FLUX 3 cannot run in ComfyUI today, because FLUX 3 has no open weights. Black Forest Labs put FLUX 3 Dev — the open-weight variant — on its roadmap without a date, and ComfyUI is a host for open-weight models: no weights, no workflow. Everything labelled "FLUX 3 ComfyUI workflow" right now is either FLUX.2 renamed or a download link for something that does not exist.
The useful version of this article is what does run today — and it is excellent: ComfyUI natively supports the FLUX.2 family, including the Apache 2.0 klein variants you can use commercially for free. This guide sets up that stack end to end, flags the licence red lines, and tells you what to prepare so you are ready the day FLUX 3 Dev ships.
Sourcing note: FLUX 3 availability, FLUX.2 specs, VRAM figures, and licences come from Black Forest Labs' official announcement, model pages, and licensing page. ComfyUI's native model support list comes from the official ComfyUI repository README. Hugging Face links point to BFL's official model cards. Last verified August 13, 2026.
What this guide solves
The pain point: ComfyUI users follow "FLUX 3" tutorials, download a workflow, and hit a wall — either the checkpoint does not exist or the graph loads a model it does not recognise. The tutorials never say the one thing that would save the hour: FLUX 3 weights are not published.
The differentiator: this guide gives the status answer first, then the working setup — which models ComfyUI officially supports, where the files go, what VRAM each needs, and which licence applies to each — plus how to verify a "FLUX 3 workflow" file before you trust it.
The straight answer on FLUX 3
Status as of August 13, 2026:
| Question | Answer |
|---|---|
| Are FLUX 3 open weights published? | No |
| Is FLUX 3 Dev announced? | Yes — on the roadmap, no date |
| Can ComfyUI run FLUX 3? | Not until weights exist |
| Does ComfyUI support FLUX.2? | Yes — natively, per the official README |
FLUX 3 Video is in request-gated early access through BFL's own channel, and FLUX 3 Image early access had not opened at last check. Neither is a download. Until BFL releases FLUX 3 Dev checkpoints on Hugging Face or GitHub, no "FLUX 3 checkpoint" file is authentic — treat every one you see as renamed FLUX.2 or malware bait.
What runs in ComfyUI today
Per the official ComfyUI README, native model support includes the FLUX line: Flux.1, Flux.2, Flux Kontext (image editing) and Flux.2 Klein. That is the entire usable FLUX set for ComfyUI right now:
| Model | What it is | Licence |
|---|---|---|
| FLUX.2 [klein] 4B / 4B Base | Smallest, fastest, ~8.4–9.2 GB VRAM | Apache 2.0 — free commercial |
| FLUX.2 [klein] 9B / 9B Base | Quality step up, ~19.6–21.7 GB VRAM | FLUX Non-Commercial licence |
| FLUX.2 [dev] | 32B, most capable open model | Non-commercial + commercial tiers |
| FLUX.1 [schnell] / [dev] | Previous generation | Apache 2.0 / Non-commercial |
For a first ComfyUI build, FLUX.2 [klein] 4B is the sane choice: Apache 2.0, commercial use included, and the lightest VRAM footprint of the family.
Setting up FLUX.2 in ComfyUI
1. Install ComfyUI. The official desktop app for Windows and macOS is the easiest route; the GitHub repository covers portable and manual installs for Linux and other GPUs.
2. Download the model files. BFL publishes FLUX.2 checkpoints under the official black-forest-labs organisation on Hugging Face. Get the model, the text encoders, and the VAE from the model card — large models are usually split across several files, and missing one fails silently.
3. Place the files where ComfyUI looks. The README's rule of thumb: checkpoints go in models/checkpoints, and models with multiple files follow each model card's instructions. When in doubt, extra model paths can be configured in extra_model_paths.yaml.
4. Load or build the graph. ComfyUI ships template workflows for the models it supports; for FLUX.2 the standard text-to-image graph covers prompt encoding, sampling, and VAE decode. Load a FLUX template and swap in your checkpoint.
5. Verify VRAM before queueing. With the official figures in front of you: klein 4B needs roughly 8.4 GB, 9B roughly 19.6 GB. If your card is below that, the queue will error or offload painfully — pick the 4B line first.
6. For maximum quality on bigger cards: FLUX.2 [dev] at 32B is documented with an fp8 path built by BFL together with NVIDIA and ComfyUI — the fp8 reference implementation is the route BFL itself points consumer GPUs toward. Check the model card for the exact files.
The licence red lines
This is the part that gets people in trouble after the setup works:
- klein 4B / 4B Base are Apache 2.0 — commercial work, client deliverables, your own fine-tunes, no permission needed.
- klein 9B and FLUX.2 [dev] are non-commercial by default. Selling outputs made with them, or building a product on them, requires a commercial licence through BFL's tiers (Builder / Platform / Professional / Enterprise).
- Fine-tuning is a separate question. BFL's licence tiers explicitly include fine-tuning and LoRA rights at specific levels; check the tier before training on a non-commercial model.
- FLUX.1 [schnell] is also Apache 2.0 — the older but still permissive fallback.
The practical rule: if you are doing commercial work on a consumer GPU today, klein 4B is the model that needs no licence conversation at all.
Want FLUX output without the VRAM spreadsheet? Flux 3 AI is an independent browser workspace running today's FLUX family — no checkpoints, no VRAM planning, no licence homework. Open the image generator or see the credit plans.
The "FLUX 3 workflow" problem
Every week a new "FLUX 3 ComfyUI workflow" surfaces. Here is how to judge any file or tutorial you find:
- Checkpoints are public or they are not. Search for FLUX 3 weights on BFL's official channels (Hugging Face organisation, GitHub). If they are not there, the workflow's checkpoint is fake or mislabelled.
- Trust the model page. BFL's FLUX 3 page lists access status; a workflow claiming to run FLUX 3 while the page says "early access by request" is describing something else.
- Read the README. ComfyUI's official support list names the models it actually runs — anything outside it needs a custom-node author you can verify.
The pattern is consistent: the files that exist are FLUX.2 checkpoints renamed, or community fine-tunes of FLUX.2 sold as "FLUX 3." The real FLUX 3 will be unmistakeable when it arrives — published by BFL, on official channels, with the roadmap date beside it.
How to prepare for FLUX 3 Dev
When FLUX 3 Dev ships, ComfyUI adoption will be fast but not instant. What you can do now:
- Build on FLUX.2 klein 4B so your ComfyUI setup, node versions, and workflow habits are current.
- Keep VRAM headroom in mind. FLUX.2 [dev] at 32B already pushes consumer GPUs to the fp8 path; a multimodal FLUX 3 Dev will not be smaller. If you are shopping hardware for FLUX 3, that is the constraint to plan around.
- Learn the prompt rules now. The official FLUX.2 prompting guide — subject first, positive phrasing, no negative prompts — is the documented family behaviour that a future FLUX 3 workflow will inherit.
- Watch the roadmap, not the tutorials. BFL's announcements are the only reliable signal for FLUX 3 Dev timing.
FAQ
Can I run FLUX 3 in ComfyUI? Not yet. FLUX 3 open weights are unpublished; FLUX 3 Dev is on BFL's roadmap with no date. ComfyUI can only run models whose weights exist.
Is there a real FLUX 3 ComfyUI workflow? No authentic one exists, because the checkpoint does not exist. Anything labelled FLUX 3 today is renamed FLUX.2 or mislabelled.
What is the best FLUX model for ComfyUI right now? FLUX.2 [klein] 4B — Apache 2.0, commercial use free, ~8.4 GB VRAM, natively supported.
How much VRAM does FLUX.2 need in ComfyUI? Official figures: klein 4B ~8.4 GB, 4B Base ~9.2 GB, klein 9B ~19.6 GB, 9B Base ~21.7 GB. FLUX.2 [dev] 32B has an fp8 path for consumer GPUs.
Can I use FLUX.2 commercially in ComfyUI? klein 4B / 4B Base are Apache 2.0 — yes, without a licence. klein 9B and dev are non-commercial by default; commercial use needs a BFL licence tier.
Where do I download FLUX.2 weights? From BFL's official Hugging Face organisation and GitHub repository. Avoid mirror links.
What will FLUX 3 Dev require? Unknown, but based on the family trajectory, expect a large multimodal checkpoint — plan your VRAM headroom now.
Bottom line
FLUX 3 in ComfyUI is a question with a waiting-room answer: the model is not out, and no workflow file can fix that. What ComfyUI runs superbly today is the FLUX.2 family — natively supported, with an Apache 2.0 klein 4B that is free for commercial work and light on VRAM.
Set that up, learn the official prompting rules, and keep an eye on BFL's roadmap. When FLUX 3 Dev ships, you will be the person who just swaps a checkpoint — not the one downloading fake workflows. Start with the image generator if you want FLUX output while you build, or compare every FLUX tier before you commit to local hardware.
Sources
- FLUX 3 — Real World Models (BFL, July 23, 2026) — FLUX 3 status, early access gating, FLUX 3 Dev roadmap
- ComfyUI official repository (README) — native model support list (Flux.1, Flux.2, Flux Kontext, Flux.2 Klein), installation routes, model folders
- FLUX.2 klein model page (BFL) — variant table, VRAM, licences, inference figures
- FLUX.2: Frontier Visual Intelligence (BFL, November 25, 2025) — FLUX.2 [dev] 32B and the fp8 reference implementation built with NVIDIA and ComfyUI
- FLUX open weights licensing (BFL) — Apache 2.0 tiers and commercial licence structure
- black-forest-labs/flux on GitHub — official open-weight releases and model list
Scope note: availability and hardware figures as of August 13, 2026 — verify at bfl.ai and the ComfyUI repository. Flux 3 AI is an independent creator workspace, not affiliated with Black Forest Labs or ComfyUI.
