Best image models for 12 GB of VRAM
Which open-weight image models run on an NVIDIA GPU with 12 GB of VRAM. Verdicts come from exact Hugging Face file sizes, model card and app maker statements, and published runs.
- Run
- 6
- Unknown
- 0
- Fits with offload
- 3
- Exceeds VRAM
- 1
6 of 10 listed image models run on an NVIDIA GPU with 12 GB of VRAM: Qwen-Image-2.1, Z-Image-Turbo, FLUX.1 [dev], FLUX 3 Image Klein, Krea 2 Turbo and SDXL Base 1.0. Highest blind-vote rank among them: Qwen-Image-2.1, #19 of 82 in text to image (Arena, snapshot 2026-10-07). FLUX.2 [klein], MOSS-Image and Stable Diffusion 3.5 Large fit with offload: the main model weights fit in VRAM; the text encoder and VAE run from system RAM, which is slower. FLUX.2 [dev] exceeds the card even for the main weights.
License check before commercial use: Qwen-Image-2.1, FLUX.1 [dev], FLUX 3 Image Klein, Krea 2 Turbo.
Get the weekly local AI newsletter
The Thursday ModelFit newsletter: new local models and what fits your machine.
For: MacBook Air M5 16 GB
You'll also get the weekly ModelFit email. By signing up you agree to our Privacy Policy. Unsubscribe from either email anytime, in one click.
Which image models run on an NVIDIA GPU with 12 GB of VRAM?
10 modelsRuns means a model card, app maker or NVIDIA figure, or a dated run on a named card, at or below 12 GB. Fits with offload means the main model weights fit in 12 GB with a 15% margin but the full set does not: the text encoder and VAE run from system RAM, which is slower. Exceeds VRAM means even the main weights are too large for the card. Mac runs never count here. unknown means no source covers this size yet; it is not a failure.
| Model | Verdict | Why | Quality | App |
|---|---|---|---|---|
| Qwen-Image-2.1Qwen Research License Agreement | Runs · Published | A dated run on RTX 5070 Ti Laptop with 12 GB of VRAM: stable-diffusion.cpp Q6_K, offload to CPU, 1024x1024.stable-diffusion.cpp issue #2118 | #19 of 82 in text to image11,441 votes | Diffusers, stable-diffusion.cpp |
| Z-Image-TurboApache 2.0 | Runs · Stated | Stated by the source: 4 GB of VRAM runs it, below this 12 GB tier (stable-diffusion.cpp with GGUF weights).stable-diffusion.cpp Z-Image docs | #59 of 82 in text to image24,808 votes | Diffusers, stable-diffusion.cpp |
| FLUX.1 [dev]FLUX.1 [dev] Non-Commercial License v1.1.1 | Runs · Stated | Stated by the source: 6 GB of VRAM runs it, below this 12 GB tier (stable-diffusion.cpp with GGUF weights).stable-diffusion.cpp FLUX docs | #78 of 82 in text to image49,221 votes | Diffusers, ComfyUI |
| FLUX 3 Image KleinFLUX Kommunity License v1.0 | Runs · Stated | Stated by the source: 8 GB of VRAM runs it, below this 12 GB tier (official ComfyUI int8 files).FLUX 3 Image Klein README (Black Forest Labs) | No independent score yet | Diffusers |
| Krea 2 TurboKrea 2 Community License Agreement v.1 | Runs · Published | A dated run on RTX 4070 with 12 GB of VRAM: ComfyUI fp8, dynamic VRAM paging, 8 steps.ComfyUI issue #15365 | No independent score yet | Diffusers, stable-diffusion.cpp |
| SDXL Base 1.0CreativeML Open RAIL++-M License | Runs · Stated | Stated by the source: 8 GB of VRAM runs it, below this 12 GB tier (SDXL 1.0, no app named).SDXL 1.0 announcement (Stability AI) | No independent score yet | ComfyUI, stable-diffusion.cpp |
| FLUX.2 [klein]Apache 2.0 (4B), FLUX Non-Commercial License (9B) | Fits with offload | The main model weights of Klein 4B Q8_0 GGUF, 4.3 GB, fit in 12 GB of VRAM with a 15% margin; the full set is 12.5 GB, so the text encoder and VAE run from system RAM, which is slower. | No independent score yet | Diffusers, stable-diffusion.cpp |
| MOSS-ImageApache 2.0 | Fits with offload | The main model weights of UD-Q4_K_XL GGUF, 3.4 GB, fit in 12 GB of VRAM with a 15% margin; the full set is 13.2 GB, so the text encoder and VAE run from system RAM, which is slower. | No independent score yet | stable-diffusion.cpp |
| Stable Diffusion 3.5 LargeStability AI Community License | Fits with offload | The main model weights of Q4_0 GGUF, 4.8 GB, fit in 12 GB of VRAM with a 15% margin; the full set is 16.4 GB, so the text encoder and VAE run from system RAM, which is slower. | #81 of 82 in text to image23,396 votes | ComfyUI |
| FLUX.2 [dev]FLUX Non-Commercial License | Exceeds VRAM | Even the main model weights are too large: the smallest complete download, Q4_K_M GGUF at 34.8 GB, has 20.1 GB of them, more than 12 GB of VRAM holds with a 15% margin. | #38 of 82 in text to image77,988 votes | Diffusers, ComfyUI, stable-diffusion.cpp |
Quality: blind human votes from Arena (arena.ai), CC BY 4.0, snapshot 2026-10-07. Scores are for the Arena checkpoint, not your local quantized build.
Frequently asked questions
What is the best image model for an NVIDIA GPU with 12 GB of VRAM?
Qwen-Image-2.1 has the highest blind-vote rank among the image models that run on an NVIDIA GPU with 12 GB of VRAM: #19 of 82 in text to image, from 11,441 votes (Arena, snapshot 2026-10-07). Z-Image-Turbo, FLUX.1 [dev], FLUX 3 Image Klein, Krea 2 Turbo and SDXL Base 1.0 also run.
Which image models run on an NVIDIA GPU with 12 GB of VRAM?
Qwen-Image-2.1 (run on RTX 5070 Ti Laptop with 12 GB), Z-Image-Turbo (4 GB stated by stable-diffusion.cpp Z-Image docs), FLUX.1 [dev] (6 GB stated by stable-diffusion.cpp FLUX docs), FLUX 3 Image Klein (8 GB stated by FLUX 3 Image Klein README (Black Forest Labs)), Krea 2 Turbo (run on RTX 4070 with 12 GB) and SDXL Base 1.0 (8 GB stated by SDXL 1.0 announcement (Stability AI)).
Which image models fit 12 GB of VRAM only with offload?
FLUX.2 [klein] (Klein 4B Q8_0 GGUF: 4.3 GB of main weights in a 12.5 GB set), MOSS-Image (UD-Q4_K_XL GGUF: 3.4 GB of main weights in a 13.2 GB set) and Stable Diffusion 3.5 Large (Q4_0 GGUF: 4.8 GB of main weights in a 16.4 GB set). The main model weights fit in VRAM; the text encoder and VAE run from system RAM, which is slower.
Which image models exceed 12 GB of VRAM even for the main weights?
FLUX.2 [dev] (Q4_K_M GGUF: 20.1 GB of main weights in a 34.8 GB set). Even its main model weights are too large for the card with a 15% margin, so part of the model itself has to sit in system RAM.