Z-Image Turbo
image gen · reviewed by Stan K. · Source: data snapshot · data as of 2026-11-12
A 6B Apache 2.0 image model with sub-second generation on an H800 at 8 NFE (9 steps) — and it still runs on a 3060 12 GB via community quants.
#2 of 20 in Image Gen
ToolScore breakdown — why 92/100
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 — commercial use, fine-tunes and serving are license-clean
- ▸Sub-second generation on an H800 at 8 NFE (9 steps) — speed and Apache in one package
- ▸Also runs on our 3060 12 GB via community quants — our experience, not an official number
- ▸16 GB VRAM is an official figure, not community guesswork
- ▸A missed generation costs a re-roll, not a credit — the economics volume work runs on
Where it costs you
- ▸No AA Elo listing in this snapshot — no arena quality reference (verify pending)
- ▸The 3060 12 GB result is community-quant territory: our experience, not an official number
- ▸Local means local — no hosted API fallback in this row
Hidden traps — the stuff the pricing page doesn't say
The official figure is 16 GB. Our 12 GB run uses community quants — our experience, not an official number. Plan hardware against 16 GB, not against our quants.
Z-Image has no AA Elo listing in this snapshot — quality reference verify pending. Our quality pillar is production experience, not blind-vote measurement.
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
The value star of paid image APIs: $0.008 base and still only $0.036 at 4K — the cheapest official p
Our measured anchor: 20 steps (cfg 2.5) in ComfyUI on the 3060 box, in-pipeline quality, Apache 2.0
AA Elo 1135 at $0.02/img — top-ten quality at the lowest AA-listed price on the board; no official p
Z-Image Turbo review
In-depth analysis · last updated 2026-11-12
The top of the local board
Z-Image Turbo tops our local table at 92 with the simplest pitch in image generation: a 6B model under Apache 2.0 that generates sub-second on an H800 at 8 NFE (9 steps). Speed and Apache in one package — no revenue cap, no territory ban, no non-commercial clause waiting in the LICENSE file. In a table where license traps cap the score, that cleanliness is most of the win.
The economics — watts, not credits
There is no per-unit price here: the cost model is watts. A 170 W card draws ≈$0.03/hour at the US average rate ($0.18/kWh), and a missed generation costs a re-roll, not a credit. That last clause is the whole local argument — on a paid API a rejected generation burns the same credits as a keeper, while here the hit rate only costs electricity.
Real limits
The quota is the GPU, not a policy: no credit pools, no rate limits. The official figure is 16 GB VRAM; our 3060 12 GB runs it via community quants, which is our experience, not an official number — budget for 16 GB and treat 12 GB as a pleasant surprise.
How the pillars land
Price/Value 10: watts-only at ≈$0.03/hour. Quotas & rate limits 10: none exist — hardware is the ceiling. Generation speed 10: sub-second on an H800 at 8 NFE (9 steps). Training use & license 9.5: Apache 2.0, nothing withheld. API & weights 9.5: open weights plus a community-quant ecosystem. Feature set 8.5: our own production experience — the arena has no Z-Image listing in this snapshot (verify pending), so this pillar is experience, not measurement. Track record 8: young model, short history, no license flips on record. Ease of use 9.5: runs on consumer cards via quants. Pricing transparency 9: publishes a real official VRAM figure. Support & docs 8: model card, LICENSE file and the quant community — no vendor support channel exists, but nothing here needs one.
Verdict — 92/100, TOP LOCAL PICK
If you own a GPU, this is the local default: fastest published generation of the open board with the cleanest license. The only caveat is honesty, not capability — quality is our experience until an arena listing lands.