Home › Image & video models › FLUX.1 Kontext [dev]
FLUX.1 Kontext [dev] VRAM requirements — what is official, and what is not
Official statement: none. The editing variant carries the same official param count as [dev]; the extra memory an editing pass needs has no official figure anywhere, so we print the gap instead of guessing it.
Where we checked, and what we did not find
- Hugging Face repo card: gated — an unauthenticated GET of the raw README returns a 140-byte “Access to model … is restricted” stub (attempted 2026-10-24; stub archived in our build notes). What is not gated is the API metadata below, and no VRAM field exists in it.
- Hugging Face API metadata for this repo: parameter count (below), license, tags — and no memory/VRAM field anywhere in the payload (2026-10-24).
- black-forest-labs/flux GitHub docs README, raw file retrieved 2026-10-24:
case-insensitive grep for
VRAMandmemory→ 0 hits.
Conclusion, stated carefully: no official VRAM requirement is published for this model — not by BFL, not in the repo metadata, not in the docs → no public data (we checked, 2026-10-24).
The arithmetic we can honestly show
- Params, official repo metadata: 11,901,408,320 (huggingface.co/api/models/black-forest-labs/FLUX.1-Kontext-dev → safetensors BF16 params 11,901,408,320, retrieved 2026-10-24).
- bf16/fp16 weight floor: 11,901,408,320 × 2 bytes ≈ 22.17 GiB — derived by us from the official count; a weights-only floor, not a running footprint.
- On a nominal 24 GB card that leaves ≈ 1.83 GiB before text encoders, VAE and activations — and that budget is already generous: we take the card’s nominal GB whole and deduct nothing for OS or display driver, because no source publishes a defensible deduction (our heuristic, stated on /vram/).
- Text encoders + VAE + activations for the FLUX pipeline: no official GB figures exist → no public data (we checked, 2026-10-24).
- Repo metadata: tags
image-to-image/image-generationand cites arXiv:2506.15742 (retrieved 2026-10-24) — this is the in-context editing variant. Its param count is identical toFLUX.1-dev(11,901,408,320 each, same day). Identical counts are arithmetic-identical budgets; whether the weights are the same is not officially explained → no public data (we checked, 2026-10-24).
Measured-anchors (deliberately empty)
Run FLUX.1 Kontext [dev] at your resolution/precision, note the peak memory your driver reports, post it anywhere public — we cite reports with handle & date, and that beats any blog number including ours. None is on this page yet. Testing without buying: RunPod RTX 5090 32GB @ $0.99/h is our cheapest ≥24 GB anchor (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. Re-check 2026-10-22 (us): source table values unchanged since its stated 2026-09-29 check). Repo demand, same retrieval as the params: 342,930 downloads, 2,916 likes (HF API).
FAQ
How much VRAM does FLUX.1 Kontext [dev] need?
No official figure exists (we checked the gated HF card via API metadata, the black-forest-labs/flux GitHub docs and the repo README on 2026-10-24). Officially checkable: 11,901,408,320 params → bf16/fp16 weights alone ≈ 22.17 GiB. Text encoders, VAE and activations stack on top with no published GB figures — a field measurement, not a guess, is the only honest completion.
Can FLUX.1 Kontext [dev] run on a 24 GB card?
Nobody official answers that. By arithmetic the bf16 weight floor ≈ 22.17 GiB leaves about 1.83 GiB of a nominal 24 GB budget for encoders, VAE and activations — before any OS/driver deduction, which we deliberately do not invent (our stated heuristic, see /vram/). Measured peaks: no public data (we checked, 2026-10-24).
Can FLUX.1 Kontext [dev] run on 16 GB or less?
The weight floor alone exceeds 16 GB by ≈ 6.17 GiB, so not with unquantised bf16 weights. Quantised community builds circulate but publish no verifiable spec → we list zero unverified numbers. Ran one? Post your peak-memory report and we cite you with handle and date.
Is the licence commercial-friendly?
the repo metadata names the licence flux-1-dev-non-commercial-license (HF API cardData, retrieved 2026-10-24) — the same non-commercial family as [dev]; schnell alone is apache-2.0. Read the licence text yourself before shipping.
FLUX family & series links: FLUX.1 overview · schnell · dev · Kontext dev · FLUX vs SDXL · SDXL · Wan 2.1 (official numbers exist) · LLM pages · method & sources.