FLUX.2-klein-4B
Black Forest Labs · image gen · reviewed by Stan K. · Source: data snapshot · data as of 2026-11-12
The Apache half of the FLUX.2 klein pair — 4B params and one of the few models that publishes a real VRAM figure (~13 GB, 3090/4070 class).
#5 of 20 in Image Gen
ToolScore breakdown — why 84/100
10 criteria × 10 points. Low pillars are visible on purpose — that's the point of a score you can audit. Methodology →
What it does well
- ▸Apache 2.0 — the permissive half of the FLUX.2 klein pair
- ▸One of the few that publishes a real VRAM figure (~13 GB, 3090/4070 class)
- ▸4B params — the light side of the family
Where it costs you
- ▸The pair is split-license: the 9B sibling is FLUX Non-Commercial — one download apart from a trap
- ▸No speed figure published in this snapshot
Hidden traps — the stuff the pricing page doesn't say
FLUX.2-klein-4B is Apache 2.0; FLUX.2-klein-9B is FLUX Non-Commercial. Same family, adjacent files, opposite commercial rights — read the LICENSE of the exact file you pull.
Know the limits before you pay: read the traps above, then decide with eyes open.
Affiliate link where available — never affects the score.
- internal ToolScore: The License Traps Hiding in “Open” Image Models — nine verified traps, one US ban verified
- internal ToolScore: FLUX vs Qwen-Image on a 3060 — our production notes from the 12GB box verified
Data as of 2026-11-12. Prices and limits change monthly — report anything stale. How we verify: methodology.
Closest alternatives
Straight PAYG — $120/1M image tokens, $0.134/img at 1–2K, $0.24 at 4K — with a SynthID watermark on
The only API here that outputs real vectors ($0.08/img) — the logo/icon niche — with V4.1 Flash rast
Few-step by design and Apache 2.0 from day one — the open sibling of dev; our 3060 box runs it quant
FLUX.2-klein-4B review
In-depth analysis · last updated 2026-11-12
The Apache half of the pair
FLUX.2-klein-4B is the permissive half of BFL's klein twins: Apache 2.0, 4B params, and one of the few models in this comparison that publishes a real VRAM figure — ~13 GB, explicitly 3090/4070 class. Most of this board hides hardware requirements behind community guesswork; this one puts the number on the model card.
The economics
Local watts math: ≈$0.03/hour on a 170 W card at the US average rate, re-rolls included. The 4B size keeps generation cheap in time as well as money.
The split-license family
The trap lives one row over: FLUX.2-klein-9B — same family, double the size — is FLUX Non-Commercial, zero commercial rights without BFL. The 4B being Apache makes the pair genuinely dangerous to shop by name alone: read the LICENSE of the exact file you pull.
How the pillars land
Price/Value 10: watts only. Quotas & rate limits 10: none — the ~13 GB bar is the ceiling. Generation speed 5: the neutral 5 — no speed figure is published in this snapshot. Training use & license 9.5: Apache 2.0. API & weights 9: open weights with a published hardware class. Feature set 8: the FLUX.2 klein line — no AA Elo listing for it in this snapshot, so quality is family inference. Track record 8: current BFL generation. Ease of use 8.5: published VRAM makes hardware planning trivial. Pricing transparency 9.5: publishing a real VRAM figure is the behavior this whole table begs for. Support & docs 6.5: model card and LICENSE; no support channel.
Verdict — 84/100
A clean Apache license with honest hardware numbers — the 4B is what you grab when you want FLUX.2-line output without reading case law first. The 5 in speed is an explicit “no data”, not a weakness.