FLUX.2 [klein] 4B on 6 GB of VRAM
Yes, but only just: FP8 leaves 0.1 GB free on 6 GB. The smallest FLUX.2: 4B parameters, Apache-2.0 licence, a few steps per image. Fits almost any modern card.
FP8 (4.1 GB) fits with only 0.1 GB to spare (about 5.9 GB of 6 GB).
Got an RTX 40 or 50? Swap Q8_0 for FP8 (4.1 GB): same quality class, computed natively there.
01What to download
| Part | File | Folder | Size | |
|---|---|---|---|---|
| Model | flux-2-klein-4b-fp8.safetensors FP8 · black-forest-labs/FLUX.2-klein-4b-fp8 | models/diffusion_models | 4.1 GB | Download → |
| Text encoder | qwen_3_4b_fp8_mixed.safetensors Qwen3 4B FP8 · Comfy-Org/z_image_turbo | models/text_encoders | 5.6 GB | Download → |
| VAE | flux2-vae.safetensors FLUX.2 VAE · Comfy-Org/flux2-dev | models/vae | 0.3 GB | Download → |
| Total download · keep about the same free on disk | 10.0 GB | |||
Alternatives: Qwen3 4B BF16 (8.0 GB). Sizes from Hugging Face, 2026-10-01.
Free ComfyUI workflow, tested on my RTX 5060 Ti: FP8 workflow (.json). Drag the file into ComfyUI. All workflows →
02Every file, on 6 GB
| File | Size | Needed | On this card | Quality | Download |
|---|---|---|---|---|---|
| 16-bit SAFETENSORS · black-forest-labs | 7.8 GB | 9.6 GB | 3.6 GB too big | the original weights | Hugging Face → |
| Q8_0 GGUF · unsloth | 4.3 GB | 6.1 GB | spills 0.1 GB | practically identical to the original | Hugging Face → |
| FP8 ← SAFETENSORS · black-forest-labs | 4.1 GB | 5.9 GB | fits · 0.1 GB spare | practically identical to the original | Hugging Face → |
| Q6_K GGUF · unsloth | 3.4 GB | 5.2 GB | fits · 0.8 GB spare | very close to the original | Hugging Face → |
| Q5_K_M GGUF · unsloth | 3.1 GB | 4.9 GB | fits · 1.1 GB spare | close; small differences in fine detail | Hugging Face → |
| Q4_K_M GGUF · unsloth | 2.6 GB | 4.4 GB | fits · 1.6 GB spare | good; some loss in fine detail and text | Hugging Face → |
| Q3_K_M GGUF · unsloth | 2.1 GB | 3.9 GB | fits · 2.1 GB spare | noticeable loss of detail | Hugging Face → |
| Q2_K GGUF · unsloth | 1.8 GB | 3.6 GB | fits · 2.4 GB spare | heavy loss; a last resort | Hugging Face → |
“Needed” adds 1 GB of working memory for FLUX.2 klein 4B and a 0.8 GB reserve. Sizes 2026-10-01.
03The text encoder
At 5.6 GB even the smallest Qwen3 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 8.0 GB, FP8 5.6 GB. Why the encoder does not count against VRAM →