TL;DR: FLUX 3 is four products, and as of October 5, 2026 exactly one has open weights — the 7B Action robotics model, which most home labs don’t need. The one you want, FLUX 3 Image, launched October 1 as API-only with open weights promised “in the coming weeks.” Don’t buy a GPU for FLUX 3 until a file size exists.

	FLUX 3 Image	FLUX 3 Video	FLUX 3 Action	FLUX 3 Dev
Best for	Text-to-image, pixel-locked editing	20s video with native audio	Robot control	The full multimodal backbone, locally
Status (Oct 2026)	API only; open weights “coming weeks”	API / partner nodes only	Open weights on Hugging Face	Announced, no date
Local VRAM today	None runnable	None runnable	7B — fits consumer cards	Unknown — no size published

Honest take: Nothing about FLUX 3 justifies a GPU purchase this week. If you have a 24GB card you’re already in the best possible seat: it runs FLUX.2 dev at Q4 today and it’s first in line when the Image weights drop. If you were going to buy anyway, a used RTX 3090 is still the entry ticket — but wait the few weeks if FLUX 3 is the only reason.

Black Forest Labs announced FLUX 3 on July 23, 2026 as something genuinely different from FLUX.1 and FLUX.2: one set of weights trained jointly on images, video, audio, and robot actions, using a flow-matching recipe the team calls Self-Flow. The pitch — from the model overview and BFL’s own materials — is that all four modalities are projections of the same physical reality, so training them together forces the model to learn physics: the sound has to match the impact, the motion has to obey the mass.

That’s the research story. The home-lab story is simpler and less romantic: can you run it, on what, and when? The answer has changed three times since July, most recently four days ago, so here is the verified state of things — and what it means for your hardware budget.

The release tracker: what actually shipped

BFL has been rolling FLUX 3 out one modality at a time. Every date below is from the launch coverage, not the announcement roadmap:

Date	Release	Access	Open weights?
Jul 23, 2026	FLUX 3 announced	Invite-only preview	No
Aug 4, 2026	FLUX 3 Video GA	API, ComfyUI partner node, fal, Pixazo	No
Sep 23, 2026	FLUX 3 Action	Hugging Face	Yes — 7B, FLUX Kommunity License v1.0
Oct 1, 2026	FLUX 3 Image	BFL API, OpenRouter, fal, Krea, Higgsfield	Promised “coming weeks”; commercial weights sold to enterprises now
TBD	FLUX 3 Dev (multimodal backbone)	—	Promised for 2026, no date, no license, no parameter count

Three things in that table matter more than the rest.

First, FLUX 3 Action is the only open-weight FLUX 3 release so far — a 7-billion-parameter “world action model” that takes camera frames, robot state, and a text instruction and returns the next 32 motor commands (about two seconds of motion) plus a predicted video of the scene. It tops NVIDIA’s RoboLab-120 benchmark at 42.2%, beating Cosmos 3 Nano’s 36.8% with less than half the parameters. The weights sit on Hugging Face under the FLUX Kommunity License v1.0 (the inference code is Apache 2.0). At 7B, the BF16 weights work out to roughly 14GB by straight arithmetic — a 16GB card holds it, a 24GB card holds it comfortably. Unless you have a robot arm, this isn’t your model, but it’s the first physical evidence of what FLUX 3 weights look like outside BFL’s cluster — and proof the open-weight promise isn’t vapor.

Second, FLUX 3 Image’s open-weight clock started October 1. BFL’s launch announcement says it plainly: commercial weights are available now for companies running image generation at scale, and “an Open Weights version of FLUX 3 Image is launching in the coming weeks.” BFL has kept this exact promise twice before — FLUX.1 dev and FLUX.2 dev both followed their API siblings — so “coming weeks” from this lab has a track record. What we don’t have is a parameter count or a file size, which is the entire ballgame for home hardware.

Third, FLUX 3 Dev — the full multimodal backbone — has no date at all. Per tracking of the Hugging Face situation, there’s no repo, no license, and no published size. If your plan was “run the video model at home,” that plan has no timeline. Treat Dev as a 2027 problem until BFL says otherwise.

What it costs to use FLUX 3 today (and why that’s your benchmark budget)

Until weights land, FLUX 3 is a metered service. Verified pricing as of October 2026:

FLUX 3 Image (per-image API tiers): $0.041 at 768², $0.048 at 1K (the default), $0.07 at 1.5K, $0.10 at 2K, and $0.607 at 4K — with a 50% launch discount running until 15:00 UTC on October 8. The endpoint is POST https://api.bfl.ai/v1/flux-3-image, and it’s also live on OpenRouter and fal if you’d rather not add another API account.

FLUX 3 Video (per-second pricing): draft mode at $0.06/second for text-to-video, full quality at $0.17/second for HD and $0.29/second for Full HD, with video-to-video at $0.41–$0.53/second. Audio is generated in the same pass at no extra cost, up to 20 seconds per clip at 720p or 1080p.

Do the math on a realistic evaluation session: 100 test images at the default 1K tier costs $4.80 ($2.40 during the launch window). A single 10-second full-quality HD video costs $1.70. That’s the honest counterweight to any GPU purchase this month — you can thoroughly evaluate whether FLUX 3 Image beats your current FLUX.2 pipeline for about the price of a coffee, before committing $1,150+ to hardware. The editing capability is the actual headline: multi-turn edits that leave every unedited pixel numerically unchanged, layout control via JSON bounding boxes, up to 10 reference images, and native output up to 4K.

The VRAM question: sizing a model that has no size

BFL has published no parameter count for FLUX 3 Image or Dev — unusual for a lab that documented every prior release in detail. So any article giving you “FLUX 3 needs X GB of VRAM” is making it up. What we can do honestly is bracket the problem with the two data points that exist.

The floor: FLUX 3 Action is 7B. BF16 weights at that size are ~14GB of VRAM before activations; FP8 halves it. If the Image model were in this class, 16GB cards would be fine. It almost certainly isn’t — Action is a specialist distill of the backbone, not the flagship.

The realistic anchor: FLUX.2 dev is 32B. It’s the most recent BFL image flagship with published numbers, and we measured the whole GGUF ladder when we self-hosted it: ~64GB in BF16, ~35GB as the FP8 file, and 19–20GB at Q4_K_S/Q4_K_M — which is exactly why a 24GB card is the real-world entry point, with the Mistral text encoder offloaded to system RAM. FLUX 3 Image replaces FLUX.2 as the image flagship and adds a layout system and 10-reference conditioning on top. The safe planning assumption is FLUX.2-class or larger:

If FLUX 3 Image ships at	BF16	FP8-class	Q4 GGUF-class	Comfortable card
~7B (Action-sized — unlikely)	~14GB	~7GB	~4–5GB	Any 12GB card
~32B (FLUX.2-sized — the anchor)	~64GB	~32–35GB	~19–20GB	24GB: used RTX 3090 / RTX 4090
Larger (multimodal tax)	80GB+	40GB+	24GB+	RTX 5090 32GB or dual-GPU

Every row below the first is an extrapolation, clearly labeled as such — the point of the table is the shape of the answer: 24GB is the credible floor for running the open-weight FLUX 3 Image at a usable quant, and 32GB+ only starts mattering if the model comes in meaningfully heavier than FLUX.2. Run your own numbers against your card with the VRAM calculator once the real file sizes post.

One more planning note from the FLUX.2 experience: the system RAM floor moved to 32GB minimum (64GB recommended) when BFL switched to an LLM text encoder, and DDR5 is the expensive line item now — a 32GB DDR5 kit runs $399–$479 as of September 2026. If FLUX 3 keeps an LLM encoder (likely), budget RAM accordingly rather than assuming it’s a rounding error.

ComfyUI readiness: API nodes now, local nodes when weights exist

As of October 5, 2026, there is no local FLUX 3 node in ComfyUI, because there are no image/video weights to load. What exists:

• FLUX 3 Video partner node — official API node covering text-to-video, image-to-video, and video-to-video. Inference runs on BFL’s servers; your GPU does nothing. You’re paying the per-second rates above through Comfy credits.
• FLUX 3 Image via API platforms — OpenRouter, fal, and BFL’s own playground. Same deal: cloud inference, metered.

When the Image weights drop, the migration path is the one FLUX.2 already paved: Comfy-Org ships repackaged safetensors, city96/Unsloth publish GGUFs within days, and your existing FLUX.2 workflow swaps the transformer loader. If you’re new to that stack, build and debug it on FLUX.2 now — the skills transfer. The first error you’ll hit on a 24GB card is the same one FLUX.2 taught us:

torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 2.50 GiB.

The fix order that works, from our FLUX.2 guide: swap the BF16 text encoder for the FP8 build, launch ComfyUI with --lowvram, and if you’re still over, drop one GGUF tier — Q4_K_M → Q4_K_S buys you about a gigabyte. Watch what’s actually happening while you generate with:

nvidia-smi --query-gpu=memory.used --format=csv -l 1

On a healthy 24GB card with a Q4_K_M-class 20GB transformer resident, you’ll see memory.used settle around 21–23GB during sampling — pinned near the ceiling is normal; climbing past it is when the OOM lands.

If image generation speed on your current card is the actual pain, NVFP4 on RTX 50-series is a today-problem with a today-solution — don’t wait for FLUX 3 to fix throughput.

What to actually buy

Prices as of September–October 2026, verified against current street prices (not MSRPs — the gap is 2× on some cards):

Your situation	The machine	Price	Where
You already own a 24GB card	Nothing. You’re ready for FLUX.2 today and FLUX 3 Image at Q4 when it ships	$0	—
Buying the entry ticket: FLUX.2 dev Q4 now, FLUX 3 Image later	Used RTX 3090 24GB	$1,150–$1,350	Check price
Same tier, want speed + warranty headroom	Used RTX 4090 24GB	$2,150–$2,350	Check price
Want 32GB insurance in case FLUX 3 is heavier than FLUX.2	RTX 5090 32GB	$3,822–$5,000 (MSRP $1,999 — nobody pays it)	Check price
Undecided — test the workload before buying anything	Rented RTX 3090 from $0.07/hr, 4090 from $0.14/hr	pay per hour	Vast.ai

The row we’d pick: none of them this week, unless the 3090 was already in your cart for FLUX.2 reasons. The open-weight drop is weeks away by BFL’s own statement, the file sizes will tell you exactly which tier you need, and an impulse RTX 5090 at $3,822+ to hedge an unknown is $2,500 of insurance against a scenario (FLUX 3 Image dramatically outgrowing FLUX.2) that BFL’s distill history argues against — every FLUX generation has shipped a Klein-class variant for smaller cards within two months of the dev release.

FAQ

Can I run any part of FLUX 3 locally right now? Only FLUX 3 Action, the 7B robotics model (Hugging Face, FLUX Kommunity License v1.0). FLUX 3 Image and Video are API-only as of October 5, 2026. For local image generation today, FLUX.2 dev is the current ceiling.

When do FLUX 3 Image open weights come out? BFL said “in the coming weeks” in the October 1 launch announcement. No exact date, no license named yet. Watch the BFL Hugging Face org rather than news aggregators — weights have historically appeared there first.

Will it run on my 16GB card? Unknown until the size publishes. If FLUX 3 Image lands FLUX.2-sized (32B), 16GB means a heavily degraded Q2/Q3 quant — the honest answer at that tier will likely be a Klein-class distill, the same way FLUX.2 Klein 4B/9B serves 8–16GB cards now.

Is the API version worth using in ComfyUI meanwhile? If you need the pixel-locked editing or 10-reference composition, yes — $0.048 per 1K image is cheap evaluation. For routine generation, local FLUX.2 or FLUX.1 Kontext costs you electricity, not per-image fees.

What about licenses for commercial work? Expect the FLUX pattern: the flagship open weights non-commercial, a smaller variant more permissive. FLUX 3 Action already set the precedent (Kommunity License weights, Apache 2.0 code). Our sister site aifoss.dev tracks the open-source imaging stack and will break down the license when the Image weights ship.

Sources

• Black Forest Labs: Introducing FLUX 3 Image (official announcement) — X
• Black Forest Labs launches FLUX 3, images and 20-second video with audio, in limited release — VentureBeat
• FLUX 3 Image launches with multi-step editing that leaves the rest of your picture alone — The Decoder
• FLUX 3 Image specs, pricing, benchmarks — Kingy AI
• FLUX 3 Video GA: 20-second native-audio video, per-second pricing — Digital Applied
• Black Forest Labs open-weights FLUX 3 Action, a 7B robotics model — AI Weekly
• FLUX 3 model overview: multimodal flow models and Self-Flow — Hugging Face community blog
• FLUX 3 and Hugging Face: when will the open-weight Dev model drop — wan27
• FLUX 3 Video partner node tutorial — ComfyUI Docs
• FLUX 3 GPU requirements and FLUX.2 sizing proxy — Spheron Network
• FLUX.3 Image API pricing — OpenRouter
• FLUX.2-dev GGUF file sizes — Unsloth, Hugging Face

Last updated October 5, 2026. Prices and specs change; verify current rates before purchasing. Release status of FLUX 3 open weights was verified against sources above on October 5, 2026 — check BFL’s channels for anything newer.

Recommended Gear

• RTX 3090 24GB (used) — the entry ticket: runs FLUX.2 dev Q4 today, first in line for FLUX 3 Image weights
• RTX 4090 24GB (used) — same VRAM tier, roughly 2× the generation speed
• RTX 5090 32GB — only if FLUX 3 ships heavier than FLUX.2 and you can’t wait to find out

Get the numbers before you buy

New GPU and mini-PC benchmarks, VRAM thresholds, and price checks — sent when there's something worth acting on, not on a schedule. No spam, unsubscribe anytime.

Email address Subscribe