FLUX.2 [dev] is an image generation model by Black Forest Labs (32B; text encoder: Mistral Small 3.1（24B）). One of the largest open models; on RTX cards you need FP8 or quantized builds.
Change precision, resolution →
| Precision | Model size | 1024×1024 | 1536×1536 |
|---|---|---|---|
| FP8 (half size, near-lossless) | 32.6 GB | 34.0 GB all in VRAM 59.0 GB | 35.0 GB all in VRAM 60.0 GB |
| GGUF Q8_0 (similar to FP8) | 34.2 GB | 35.6 GB all in VRAM 60.6 GB | 36.6 GB all in VRAM 61.6 GB |
| GGUF Q5_K_S | 22.7 GB | 25.0 GB all in VRAM 49.1 GB | 25.1 GB all in VRAM 50.1 GB |
| GGUF Q4_K_S (popular low-VRAM choice) | 19.2 GB | 25.0 GB all in VRAM 45.5 GB | 25.0 GB all in VRAM 46.6 GB |
| GGUF Q3_K_S (lower quality) | 15.0 GB | 25.0 GB all in VRAM 41.4 GB | 25.0 GB all in VRAM 42.4 GB |
Bold = VRAM needed when ComfyUI swaps the text encoder and the model. "All in VRAM" = keeping both loaded. Text encoder assumed at FP8. Official: on consumer RTX GPUs, use the optimized FP8 build (measured with the official implementation).
| GPU | VRAM | Best precision | VRAM needed | Verdict | Buy |
|---|---|---|---|---|---|
| RTX 5090 | 32 GB | GGUF Q5_K_S | 25.0 GB | ○ Runs with model swapping | Check price |
ComfyUI also runs on AMD Radeon (ROCm) and Intel Arc, but setup is harder than on NVIDIA and some features (such as FP8 acceleration) are not supported.
→ Image & Video AI VRAM Calculator (all 15 models) / Local LLM VRAM Calculator
Estimates for ComfyUI, about ±20%. Model specs from the publisher's config files and released file sizes (checked October 2026).