TL;DR: FLUX.2 [dev] is a 32B-parameter image model that runs locally in ComfyUI, but only the small FLUX.2 [klein] 4B variant is Apache 2.0 — dev and klein 9B ship under Black Forest Labs’ non-commercial license. A 24GB GPU runs dev at GGUF Q4 (~19GB); 8GB cards should run klein 4B instead.

	FLUX.2 [dev] 32B	FLUX.2 [klein] 9B	FLUX.2 [klein] 4B
Best for	Max quality on 24GB+ GPUs	Mid-range quality, 16GB cards	8-12GB cards, commercial use
License	FLUX Non-Commercial v2.1	FLUX Non-Commercial	Apache 2.0
VRAM (practical)	~19GB at GGUF Q4, ~32GB FP8	~12-16GB	~8GB

Honest take: If you need commercially safe weights, klein 4B is the only Flux 2 you should self-host. If it’s personal work on a 24GB card, dev at Q4 or FP8 is the best local image model you can run in September 2026.

What license does Flux 2 ship under?

Only one Flux 2 variant is genuinely open source. Per the official FLUX.2 repository (checked September 29, 2026):

Model	Released	License
FLUX.2 [dev] (32B)	Nov 25, 2025	FLUX Non-Commercial License v2.1
FLUX.2 [klein] 9B / 9B Base / 9B KV	Jan 15, 2026	FLUX Non-Commercial License
FLUX.2 [klein] 4B / 4B Base	Jan 15, 2026	Apache 2.0

The FLUX Non-Commercial License v2.1 allows non-commercial, non-production use of the weights. Generated outputs can be used commercially — Black Forest Labs states this explicitly on the FLUX.2-dev model card — but running the model itself inside a commercial product or paid service requires their separate Self-Hosted Commercial License. This mirrors FLUX.1: dev was non-commercial there too, and only FLUX.1 [schnell] was Apache 2.0.

Two practical consequences for self-hosters:

• Freelancers and studios: generating client images with dev sits in a gray zone — the outputs are fine, but “non-production use” of the weights is hard to square with billable work. Klein 4B (Apache 2.0) removes the question entirely.
• The Hugging Face repo is gated. You must log in and accept the license before any download works. This bites during setup (fix below).

What changed from FLUX.1 to FLUX.2?

FLUX.2 [dev] jumps from 12B to 32B parameters and swaps the entire text-encoding stack. FLUX.1 used CLIP + T5-XXL; FLUX.2 uses Mistral 3 Small as its text encoder, which handles long, structured prompts and in-image text noticeably better. It also ships a new dedicated Flux 2 VAE, so none of your FLUX.1 encoder or VAE files carry over.

The size jump is the story for self-hosters. FLUX.1 dev at FP8 fit comfortably in 12-16GB. FLUX.2 dev at full BF16 wants ~64GB — H100 territory — so every consumer-GPU deployment of dev runs a compressed variant: FP8 (~32GB) or GGUF quantizations. That’s why Black Forest Labs shipped the klein distillations in January 2026: klein 4B fits in ~8GB VRAM per the official README.

How much VRAM does FLUX.2 dev need?

About 19GB at GGUF Q4_K_M, ~32GB at FP8, and ~64GB at BF16 (community-verified numbers from WillItRunAI and the ComfyUI GGUF guides, September 2026). Mapped to real GPUs:

Your GPU	VRAM	What to run
RTX 5090 / used RTX 4090	32 / 24GB	dev FP8 (5090) or dev GGUF Q4-Q5 (4090)
Used RTX 3090	24GB	dev GGUF Q4_K_M (~19GB)
RTX 5060 Ti 16GB / 4060 Ti 16GB	16GB	klein 9B, or dev at Q2-Q3 (visible quality loss)
RTX 3060 12GB	12GB	klein 4B, or klein 9B GGUF
8GB cards (RTX 4060, 3070)	8GB	klein 4B only

Three caveats:

• The Mistral 3 Small text encoder needs memory on top of the diffusion model. The FP8 encoder file is ~4GB; ComfyUI offloads it to system RAM between stages, so 32GB+ system RAM keeps that swap from thrashing. With DDR5 64GB kits at $680-$1,070 (September 2026 street prices), 32GB is the sane floor — don’t buy RAM you don’t need this year.
• Below Q4, Flux models lose fine text rendering first. If your use case is posters or UI mockups with legible text, Q2 defeats the purpose — run klein 9B at a higher quant instead.
• These numbers are for 1024×1024. 2K generations push VAE decode over the top on 12GB cards; enable tiled VAE decode if you see out-of-memory at the decode step.

How do you set up FLUX.2 in ComfyUI?

ComfyUI supports FLUX.2 natively — no custom nodes needed for the FP8 path. Update to the latest release first (git pull in your ComfyUI directory, or use the desktop app updater), because Flux 2 nodes landed after the November 2025 builds.

Download three files after accepting the license on Hugging Face:

1. Diffusion model → ComfyUI/models/diffusion_models/ — flux2_dev_fp8 (or a GGUF from the community quant repos)
2. Text encoder → ComfyUI/models/text_encoders/ — mistral_3_small_flux2_fp8.safetensors
3. VAE → ComfyUI/models/vae/ — flux2-vae.safetensors

The first problem you’ll actually hit: FLUX.2-dev is a gated repo, and an anonymous download fails with a 401 even though the URL looks public:

$ huggingface-cli download black-forest-labs/FLUX.2-dev
401 Client Error: Unauthorized for url:
https://huggingface.co/black-forest-labs/FLUX.2-dev/resolve/main/...
Access to model black-forest-labs/FLUX.2-dev is restricted.

Fix: open the model page in a browser, accept the license terms while logged in, then authenticate the CLI with huggingface-cli login (paste a read-scoped token from your HF settings). The download works immediately after. The Comfy-Org repackaged files and community GGUFs are not gated, which is one reason most local users go that route.

For GGUF quants, install the ComfyUI-GGUF custom node pack and load the model through Unet Loader (GGUF) instead of the standard loader — the stock workflow otherwise stays identical: UNETLoader → CLIPLoader (type flux2, pointing at the Mistral encoder) → VAELoader. ComfyUI’s official Flux 2 example page has a drag-and-drop workflow JSON; start from that rather than rewiring a FLUX.1 graph, because the encoder and VAE nodes are not interchangeable. If you’re new to custom node management, our ComfyUI custom nodes guide covers the install and update workflow.

On a used RTX 3090 with the Q4_K_M quant, expect roughly 60-90 seconds per 1024×1024 image at 20-28 steps — usable for iteration, not for batch production.

Is upgrading from FLUX.1 worth it?

For 24GB-card owners doing personal work: yes, dev at Q4 is a real upgrade — prompt adherence and in-image typography are where FLUX.1 users will notice the difference, thanks to the Mistral encoder. For 8-16GB cards, the honest answer is murkier. Community comparisons through mid-2026 generally put klein 9B near FLUX.1 dev quality at lower VRAM, but klein 4B trades noticeable detail for its Apache 2.0 license and 8GB footprint. If you’re on a 12GB card and happy with FLUX.1 dev FP8, nothing forces the move.

Where Flux 2 fits against the wider field — SDXL’s ecosystem depth, SD 3.5’s licensing mess — is a different question; see our Flux vs SDXL vs SD 3.5 comparison. And if you’re choosing a frontend first, ComfyUI vs Automatic1111 vs Forge explains why ComfyUI is the only one with day-one Flux 2 support.

When NOT to self-host Flux 2

• You’re selling access or embedding it in a product. Dev and klein 9B weights are non-commercial; only klein 4B (Apache 2.0) or BFL’s paid commercial license covers that.
• Your GPU has 8GB or less and you want dev-level quality. No quant gets 32B into 8GB usably. Run klein 4B, or rent — a 24GB RTX 3090 on Vast.ai starts from around $0.07/hr (marketplace pricing, September 2026), so testing dev for an evening costs less than a coffee.
• You need batch throughput. At ~60-90s per image on a 3090, a 500-image job is a 10-hour run. Cloud inference APIs beat local for one-off bulk jobs.
• AMD cards below 24GB. ROCm support for the Flux 2 stack in ComfyUI still lags NVIDIA; expect slower generation and more setup friction than the equivalent GeForce card.

What to actually buy

Prices as of September 2026, matched to the VRAM table above:

Your situation	The hardware	Price	Where
Want dev at Q4 on the cheapest 24GB card	Used RTX 3090	~$1,399-$1,450	Check price
Want dev FP8 speed and headroom	Used RTX 4090	~$2,575-$3,010	Check price
New card, klein 9B territory	RTX 5060 Ti 16GB	~$679-$805 (MSRP $429)	Check price
Undecided — test dev before buying	Rented RTX 3090/4090	from $0.07/hr	Vast.ai

GPU street prices moved hard in 2026 (the RTX 5060 Ti sells 58-88% over MSRP), so treat MSRP figures in older guides as fiction — the full picture is in runaihome’s GPU buying guide for local AI.

FAQ

Can I use images generated with FLUX.2 dev commercially? Yes. The FLUX [dev] Non-Commercial License explicitly allows personal, scientific, and commercial use of generated outputs. The restriction applies to the weights: you can’t run dev inside a commercial product or paid service without Black Forest Labs’ separate commercial license.

Does FLUX.2 work with Automatic1111 or Forge? Not natively as of September 2026. ComfyUI is the reference implementation with built-in Flux 2 nodes. Forge inherited FLUX.1 support but Flux 2’s new Mistral text encoder and VAE require updated loaders — check your fork’s release notes before assuming compatibility.

Do my FLUX.1 LoRAs work with FLUX.2? No. FLUX.2 is a different architecture (32B vs 12B, different text encoder), so FLUX.1 LoRAs are incompatible. The klein models are also distinct architectures — LoRAs must be trained per model.

Sources

• black-forest-labs/flux2 — official inference repo, model list and licenses
• FLUX.2-dev model card and license — Hugging Face
• ComfyUI official Flux 2 example workflow
• FLUX.2 dev VRAM requirements — WillItRunAI
• ComfyUI-GGUF loader — city96

What self-hosting actually costs

Real cost breakdowns for self-hosted AI: hardware floors, power, maintenance hours, and the honest comparison against paying for it. No spam, unsubscribe anytime.

Email address Subscribe