HiDream-O1-Image on 10 GB of VRAM
Not entirely. Even FP8 is 0.9 GB over 10 GB, so ComfyUI keeps part of it in system RAM. A 2026 all-in-one 8.8B model built on Qwen3-VL: no separate text encoder and no VAE. It generates, edits and personalises up to 2048×2048.
Nothing fits entirely, but FP8 (8.1 GB) overflows by only about 0.9 GB. ComfyUI keeps that part in system RAM automatically: slower than a full fit, but usable.
Got an RTX 40 or 50? Swap Q8_0 for FP8 (8.1 GB): same quality class, computed natively there.
Also try the FP8 file (8.1 GB) on NVIDIA GPUs. It does not fit entirely, but a current ComfyUI streams the overflow from RAM and native 8-bit files often beat a small GGUF that does fit. Dynamic VRAM →
01Other amounts of VRAM
02Every file, on 10 GB
| File | Size | Needed | On this card | Quality | Download |
|---|---|---|---|---|---|
| 16-bit SAFETENSORS · Comfy-Org | 16.4 GB | 19.2 GB | 9.2 GB too big | the original weights | Hugging Face → |
| FP8 ← SAFETENSORS · Comfy-Org | 8.1 GB | 10.9 GB | spills 0.9 GB | practically identical to the original | Hugging Face → |
“Needed” adds 2 GB of working memory for HiDream-O1 and a 0.8 GB reserve. Sizes 2026-10-01. The Comfy-Org files are complete checkpoints (loaded from the checkpoints folder); there is no separate text encoder or VAE to add. No GGUF version was found.
03What to download
| Part | File | Folder | Size | |
|---|---|---|---|---|
| Checkpoint | hidream_o1_image_fp8_scaled.safetensors FP8 · Comfy-Org/HiDream-O1-Image | models/checkpoints | 8.1 GB | Download → |
| Total download · keep about the same free on disk | 8.1 GB | |||
One file: this model has no separate text encoder or VAE. Sizes from Hugging Face, 2026-10-01.
04The text encoder
This model has no separate text encoder: reading the prompt is built into the model itself, so its memory is already in the file sizes above.