Mage-Flow (Microsoft) on 6 GB of VRAM
Yes, but only just: INT8 leaves 0.0 GB free on 6 GB. Microsoft's 4B native-resolution model from July 2026, with separate Turbo (4-step) and Edit checkpoints of the same size. Light enough for most GPUs.
Runs well
Yes, just.
INT8 (4.2 GB) fits with only 0.0 GB to spare (about 6.0 GB of 6 GB).
Best fileINT8
File size4.2 GB
VRAM needed~6.0 GB
System RAM32 GB+
Memory map · 6 GB GPU6.0 GB / 6 GB
03 GB6 GB
Weights INT8 · 4.2 GBWorking memory · 1.0 GBReserve · 0.8 GBFree · 0.0 GB
calculated How this is counted.
01What to download
| Part | File | Folder | Size | |
|---|---|---|---|---|
| Model | mage_flow_int8_convrot.safetensors INT8 · Comfy-Org/Mage-Flow | models/diffusion_models | 4.2 GB | Download → |
| Text encoder | qwen3vl_4b_bf16.safetensors Qwen3-VL 4B BF16 · Comfy-Org/Mage-Flow | models/text_encoders | 8.9 GB | Download → |
| VAE | mage_flow_vae_bf16.safetensors Mage-Flow VAE · Comfy-Org/Mage-Flow | models/vae | 0.3 GB | Download → |
| Total download · keep about the same free on disk | 13.4 GB | |||
Sizes from Hugging Face, 2026-10-01.
02Every file, on 6 GB
| File | Size | Needed | On this card | Quality | Download |
|---|---|---|---|---|---|
| 16-bit SAFETENSORS · Comfy-Org | 8.2 GB | 10.0 GB | 4.0 GB too big | the original weights | Hugging Face → |
| INT8 ← SAFETENSORS · Comfy-Org | 4.2 GB | 6.0 GB | fits · 0.0 GB spare | practically identical to the original | Hugging Face → |
“Needed” adds 1 GB of working memory for Mage-Flow and a 0.8 GB reserve. Sizes 2026-10-01.
03The text encoder
At 5.2 GB even the smallest Qwen3-VL 4B file is too big for 6 GB, so it runs from system RAM. That slows down reading the prompt, not the picture. Sizes: 16-bit 8.9 GB, FP8 5.2 GB. Why the encoder does not count against VRAM →