Home › Image & video models › FLUX

FLUX.1 VRAM requirements — what is official, and what is not

Official statement: none. As of 2026-10-20 Black Forest Labs publishes no VRAM requirement on the model cards, the black-forest-labs/flux docs, or bfl.ai. Anyone quoting "FLUX needs exactly N GB" is quoting a guess. What is official: huggingface.co/api/models/black-forest-labs/FLUX.1-schnell → safetensors BF16 params 11,891,178,560, retrieved 2026-10-20.

The arithmetic we can honestly show

• Params: 11,891,178,560 (huggingface.co/api/models/black-forest-labs/FLUX.1-schnell → safetensors BF16 params 11,891,178,560, retrieved 2026-10-20).
• bf16 weight floor: 11.89B × 2 bytes ≈ 22.1 GiB — derived by us from the official param count; a floor, not a running footprint.
• On a 24 GB card that leaves < 2 GiB before T5-XXL/CLIP encoders and activations → bf16 realistically wants 32 GB+, but the exact number is a measurement, and we won't invent one.

Measured-anchors (deliberately empty)

Run schnell/dev at your resolution and precision, note peak memory, post it anywhere public — we cite reports with handle & date. That beats any blog number, ours included. Renting to test: RunPod RTX 5090 32GB @ $0.99/h is the cheapest ≥24 GB anchor in our table (modelvram.com public table (self-stated check 2026-09-29, retrieved 2026-10-20). Partial re-verify 2026-10-20: Lambda A100 40GB $1.99/h CONFIRMED on lambda.ai/pricing (fetched directly); Lambda H100 rows ambiguous on-page ($3.99 vs $4.09 configs) → not adopted; RunPod & Vast pricing pages are JS-rendered, no extractable static figures — verify rates visually before spending).

FAQ

How much VRAM does FLUX.1 [dev] need?

No official figure exists (we checked the gated model card, the black-forest-labs/flux GitHub docs and bfl.ai on 2026-10-20). Officially checkable: 11.89B params → bf16 weights alone ≈ 22.1 GiB. Text encoders (T5-XXL, CLIP) and activations stack on top — that part needs a field measurement, not a guess.

Can FLUX run on a 12 GB or 16 GB card?

No official answer. Quantised community builds circulate widely but publish no verifiable spec; we list zero unverified numbers. If you ran one, tell us peak memory and we'll cite you.

What does trying FLUX cost without buying hardware?

bf16 weight floor ≈ 22.1 GiB won't fit a 24 GB card with encoders included; on our cheapest cited 32 GB rental (RunPod RTX 5090 32GB @ $0.99/h, modelvram.com public table (self-stated check 2026-09-29, retrieved 2026-10-20). Partial re-verify 2026-10-20: Lambda A100 40GB $1.99/h CONFIRMED on lambda.ai/pricing (fetched directly); Lambda H100 rows ambiguous on-page ($3.99 vs $4.09 configs) → not adopted; RunPod & Vast pricing pages are JS-rendered, no extractable static figures — verify rates visually before spending) an exploratory hour costs that much.

Is FLUX.1 [schnell] free for commercial use?

schnell is apache-2.0 per the HF repo metadata (retrieved 2026-10-20); the dev weights carry a separate non-commercial licence — read the licence page on the repo before shipping.

Video-model sibling: Wan 2.1 VRAM (official numbers) · LLMs: model pages · method · embed badge →.