Thurmon Demich

Posted on Jul 29 • Originally published at bestgpuforai.com

Best GPU for FLUX 3: What to Buy Days After Launch

#flux3 #flux #imagegeneration #videogeneration

From the Best GPU for AI archive. The canonical version has interactive calculators, an up-to-date GPU comparison table, and live pricing.

Black Forest Labs shipped FLUX 3 on July 23, 2026 — a unified model that generates image, video, and audio from one set of open weights. So which GPU actually runs it? Here's the uncomfortable truth: two days after release, nobody outside BFL and their launch partners can answer that with hard numbers. What I can do is separate what's confirmed from what the Flux family's track record strongly implies, and tell you which cards are safe to buy right now without regretting it once real benchmarks land.

What's confirmed vs. what we expect

Confirmed, as of late July 2026:

FLUX 3 released July 23, 2026, with open weights.
It's genuinely multimodal — image, video, and audio generation in a single unified model, not three bolted-together checkpoints.

That's it. No official VRAM requirements table, no ComfyUI reference workflow with published memory footprints, no third-party benchmark sweep. Any article claiming exact FLUX 3 generation times this week is making them up.

Everything below is inference from how this family has behaved before — and Flux has been unusually consistent:

Flux 2 Dev wanted roughly 24GB in FP16 and around 12GB in FP8. A unified multimodal model almost certainly sits above that, because video generation carries temporal layers and much larger latent buffers than still images.
Flux 2 Klein squeezed a 4B distilled variant into about 10GB. BFL has shipped a small distilled sibling for every major release, and I'd be surprised if FLUX 3 broke the pattern.
NVIDIA shipped NVFP4 and FP8 Flux 2 variants within weeks of launch, with roughly 2.5-3x speedups and a 55-60% VRAM cut — but the NVFP4 gains were Blackwell-only.

My read: expect image-only FLUX 3 inference to land near Flux 2's footprint, and expect the video and audio paths to push well past 24GB in full precision until quantized checkpoints arrive. If you ran the numbers for Flux 2 hardware, treat those as the floor, not the target.

VRAM chart available at the original article

Expected hardware by workload

Labeled clearly as projection — these bands come from Flux 2's measured behavior plus the overhead multimodal architectures typically add:

FLUX 3 workload	Expected VRAM band	GPU pick today	Price
Image-only, quantized (FP8/NVFP4)	~12-16GB	RTX 5070 Ti 16GB	~$750
Image-only, full precision	~24-28GB	RTX 4090 24GB	~$1,600
Image + video pipeline, quantized	~20-28GB	RTX 4090 24GB	~$1,600
Image + video + audio, full pipeline	~28-40GB+	RTX 5090 32GB	~$2,000
Distilled "Klein-class" variant (if it comes)	~10-12GB	RTX 4070 Ti Super 16GB	~$700

The pattern worth internalizing: image generation is the cheap part. The moment FLUX 3's video path enters your workflow, VRAM demand jumps a tier, because you're holding dozens of frames of latents plus temporal attention state instead of one. Audio is comparatively light, but it stacks on top of everything else in a unified pipeline.

Why Blackwell NVFP4 matters more this time

When NVIDIA optimized Flux 2, the FP8 path helped every card from Ada up — but the NVFP4 checkpoints, with their 2.5-3x throughput gains and roughly 55-60% VRAM reduction, only ran on Blackwell tensor cores. If NVIDIA repeats that playbook for FLUX 3 (and a flagship open-weights multimodal release is exactly the kind of model they optimize first), the gap between RTX 40-series and 50-series widens from "somewhat faster" to "runs workloads the older card physically can't fit."

That's the strongest argument for buying Blackwell today even without FLUX 3 benchmarks. A 16GB RTX 5070 Ti with a future NVFP4 checkpoint could plausibly fit a video pipeline that a 16GB RTX 4070 Ti Super never will. I covered how this dynamic played out between the flagships in RTX 5090 vs 4090 for Flux 2 — the short version is that architecture generation started mattering as much as raw VRAM, and FLUX 3 will likely push that further.

Which GPU should you buy today?

You want zero regret and you'll use the full multimodal pipeline: RTX 5090, ~$2,000. 32GB covers the worst-case VRAM projection and Blackwell catches every NVIDIA optimization. This is the only unconditional recommendation I can make this week.
You mainly generate images and video is a curiosity: RTX 4090, ~$1,600. 24GB runs any plausible quantized FLUX 3 image workload and probably a quantized video path. You give up future NVFP4 checkpoints — go in with eyes open.
You're budget-limited but want to stay Blackwell: RTX 5070 Ti, ~$750. 16GB is my floor for FLUX 3; below that you're betting everything on a distilled variant that doesn't exist yet.
You find a discounted RTX 4070 Ti Super: Fine for image-only work at ~$700, and it runs Flux 2 beautifully today. Just accept it's probably locked out of FLUX 3's video path long-term.
You're still happily on Flux 1: No urgency. My original Flux buyer's guide picks all still hold for that model, and FLUX 3's requirements will be far clearer in a month.

There's a fifth option I'd genuinely consider right now: rent. An hour of cloud RTX 5090 or A100 time costs a few dollars, which is a cheap way to test FLUX 3's actual behavior on your workflows before committing $2,000 to a guess. Our GPU rental guide covers the cost math.

Common mistakes in week one

Buying an 8-12GB card for a multimodal model. FLUX 3's video path will not fit, quantized or not, on 8GB — and 12GB looks marginal even for images once you add control modules. 16GB is the realistic floor.
Assuming your Flux 2 workflow transfers. Every major Flux release changed the text encoder pairing, sampler recommendations, or node graph. Expect the same, and wait for official example workflows instead of debugging a Franken-graph.
Waiting indefinitely for perfect information. If your current card can't run Flux 2 comfortably, it can't run FLUX 3 either — the upgrade decision doesn't actually depend on benchmarks that haven't shipped.
Panic-selling a 4090. It's still a 24GB card that runs everything released before last week. Wait for real FLUX 3 numbers before eating a used-market loss.

Final verdict

Budget	GPU	FLUX 3 outlook (projected)
~$700	RTX 4070 Ti Super 16GB	Image-only, quantized; likely no video path
~$750	RTX 5070 Ti 16GB	Image now; NVFP4 could unlock quantized video
~$1,600	RTX 4090 24GB	Full-precision images, quantized video plausible
~$2,000	RTX 5090 32GB	Full multimodal pipeline, catches every optimization

FLUX 3 is two days old and the honest hardware advice is conditional: buy 32GB if you refuse to gamble, 16GB Blackwell if you're betting on quantization, and rent by the hour if you'd rather let the benchmarks settle first.

Frequently asked questions

How much VRAM will FLUX 3 need?

Nobody has published official requirements as of late July 2026. Based on Flux 2's measured footprint and typical multimodal overhead, expect roughly 12-16GB for quantized image-only work, somewhere past 24GB for video pipelines in higher precision, and comfortable headroom only at 32GB. Treat these as projections until real benchmarks land.

Can FLUX 3 run on an RTX 4090?

Very likely yes for image generation, and plausibly for quantized video work, given the 4090 handled Flux 2 in FP16 with room to spare. The open question is optimization: if NVIDIA ships Blackwell-only NVFP4 checkpoints as they did for Flux 2, the 4090 keeps working but misses the largest speed and VRAM gains.

Should I wait for FLUX 3 benchmarks before buying a GPU?

If your current card already struggles with Flux 2, no — anything that fails today fails harder on a larger multimodal model, so the upgrade logic is already settled. If you're choosing between two capable cards, waiting a few weeks costs little, and renting cloud GPU time in the meantime lets you test FLUX 3 on real workloads first.

Will FLUX 3 get an FP8 or NVFP4 version?

There's no announcement yet, but the precedent is strong: NVIDIA published optimized Flux 2 variants within weeks of that launch, with big throughput gains and VRAM cuts of over half on Blackwell hardware. A flagship open-weights release is exactly the model class they prioritize, so quantized checkpoints seem probable rather than guaranteed.

Related guides on Best GPU for AI

Continue on Best GPU for AI for the complete guide with interactive calculators and current GPU prices.

DEV Community