Can the Arc B580 12 GB run FLUX.1 [dev]?
Not at full quality. The 16-bit and 8-bit files are too big for the Arc B580 12 GB; the largest that fits is Q5_K_S (8.3 GB).
Q5_K_S (8.3 GB) fits with only 1.4 GB to spare (about 10.6 GB of 12 GB). Quality: close; small differences in fine detail. For better quality, Q6_K (9.9 GB) also runs, with about 0.2 GB spilling into system RAM - a little slower, still practical.
01What to download
| Part | File | Folder | Size | |
|---|---|---|---|---|
| Model | flux1-dev-Q5_K_S.gguf Q5_K_S · city96/FLUX.1-dev-gguf | models/unet | 8.3 GB | Download → |
| Text encoder | t5xxl_fp8_e4m3fn.safetensors T5-XXL FP8 · comfyanonymous/flux_text_encoders | models/text_encoders | 4.9 GB | Download → |
| Text encoder | clip_l.safetensors CLIP-L · comfyanonymous/flux_text_encoders | models/text_encoders | 0.2 GB | Download → |
| VAE | ae.safetensors FLUX.1 VAE (ae) · Comfy-Org/Lumina_Image_2.0_Repackaged | models/vae | 0.3 GB | Download → |
| Total download · keep about the same free on disk | 13.8 GB | |||
Alternatives: T5-XXL FP16 (9.8 GB). Sizes from Hugging Face, 2026-10-01.
Free ComfyUI workflow, tested on my RTX 5060 Ti: FP8 workflow (.json) · GGUF Q8_0 workflow (.json). Drag the file into ComfyUI. All workflows →
02Every FLUX.1 dev file on 12 GB
| File | Size | Needed | On this card | Quality | Download |
|---|---|---|---|---|---|
| 16-bit SAFETENSORS · black-forest-labs | 23.8 GB | 26.1 GB | 14.1 GB too big | the original weights | Hugging Face → |
| Q8_0 GGUF · city96 | 12.7 GB | 15.0 GB | 3.0 GB too big | practically identical to the original | Hugging Face → |
| FP8 SAFETENSORS · Kijai | 11.9 GB | 14.2 GB | 2.2 GB too big | practically identical to the original | Hugging Face → |
| Q6_K GGUF · city96 | 9.9 GB | 12.2 GB | spills 0.2 GB | very close to the original | Hugging Face → |
| Q5_K_S ← GGUF · city96 | 8.3 GB | 10.6 GB | fits · 1.4 GB spare | close; small differences in fine detail | Hugging Face → |
| Q4_K_S GGUF · city96 | 6.8 GB | 9.1 GB | fits · 2.9 GB spare | good; some loss in fine detail and text | Hugging Face → |
| Q3_K_S GGUF · city96 | 5.2 GB | 7.5 GB | fits · 4.5 GB spare | noticeable loss of detail | Hugging Face → |
| Q2_K GGUF · city96 | 4.0 GB | 6.3 GB | fits · 5.7 GB spare | heavy loss; a last resort | Hugging Face → |
“Needed” adds 1.5 GB of working memory for FLUX.1 dev and a 0.8 GB reserve. Sizes 2026-10-01.
03Measured and reported results
| Label | Setup | Result | Peak VRAM | Date | Source |
|---|---|---|---|---|---|
| reported | fp8 · 1024x1024 · 20 steps “Flux1 dev fp8 (20step): GOOD (35s)” Intel Arc B580 12GB, PyTorch 2.8 XPU in Docker (YanWenKun); single 1024x1024 image, pre-warmed inference only (model load excluded) | 35 s / image | — | 2025-08-03 | github.com → |
Other people's numbers are copied as published, with a link. Send yours.
04With more VRAM
With 16 GB you could run FLUX.1 dev at 8-bit or better (Q8_0, 12.7 GB) with no offloading: for example on the RTX 5080 16 GB, RTX 5070 Ti 16 GB, RTX 5060 Ti 16 GB.
What changes from the Arc B580 12 GB to the RTX 5060 Ti 16 GB →
05About the Arc B580 12 GB
Intel Arc GPUs run ComfyUI through PyTorch's native XPU support on Windows 11 and Linux. Memory works the same as on NVIDIA, so the fit verdicts apply. FP8 files save memory but give no speed-up, and some custom nodes are NVIDIA-only.
12 GB GDDR6 · 192-bit · 456 GB/s · Battlemage. Everything that runs on it →
06Questions
How much VRAM does FLUX.1 dev need?
Around 6.3 GB with the smallest file (Q2_K) and 14.2 GB with an 8-bit file (FP8). The full 16-bit file needs about 26.1 GB.
Which FLUX.1 dev file should I download for the Arc B580 12 GB?
Q5_K_S (8.3 GB) from city96/FLUX.1-dev-gguf. Load it with Unet Loader (GGUF) from the ComfyUI-GGUF node pack.
FP8 or Q8_0 for FLUX.1 dev on the Arc B580 12 GB?
Q8_0 (12.7 GB). The Arc B580 12 GB has no FP8 maths, so FP8 (11.9 GB) only saves memory and is a step further from the original.
How much do I need to download for FLUX.1 dev on the Arc B580 12 GB?
13.8 GB: flux1-dev-Q5_K_S.gguf, t5xxl_fp8_e4m3fn.safetensors, clip_l.safetensors, ae.safetensors.