For most big models you can download the same weights three ways: the original 16-bit file (BF16 or FP16), an FP8 file of about half the size, or a GGUF file in one of many sizes. Here is what actually differs.
Size
FP8 and GGUF Q8_0 are almost the same size. For FLUX.1 [dev]: FP8 is 11.9 GB, Q8_0 is 12.7 GB, the original 23.8 GB. Below 8 bits, only GGUF gives you options (Q6, Q5, Q4…).
Speed depends on your GPU
RTX 40 and RTX 50 (Ada Lovelace and Blackwell, desktop and laptop) and AMD RX 9000 (RDNA 4) have hardware FP8. ComfyUI computes FP8 natively on NVIDIA GPUs with compute capability 8.9 or higher, and on RX 9000 with PyTorch 2.7+ and ROCm 6.4+. Native FP8 maths is used for files made for ComfyUI (Comfy-Org's fp8_scaled files); plain FP8 files use it with the --fast fp8_matrix_mult launch option.
RTX 30, RTX 20, AMD RX 7000/6000 and Intel have no FP8 compute. ComfyUI still loads the FP8 file, but converts the weights back to higher precision before the maths. You save memory, not time.
RTX 50 GPUs also support FP4 (NVFP4 files), which is faster again. More in the new file formats of 2026.
Quality
Q8_0 is usually a hair closer to the original than FP8, because it stores a scale for every small block of weights. In real pictures the difference is very small.
So which one?
- RTX 40 / RTX 50 / RX 9000, and the 8-bit file fits: FP8.
- RTX 30 or older, older AMD, Intel, and the 8-bit file fits: Q8_0.
- 8-bit does not fit: on NVIDIA with an up-to-date ComfyUI, try the native FP8/INT8 file anyway — Dynamic VRAM streams the rest. Otherwise take the biggest GGUF that fits (Q6_K, then Q5, then Q4).
The verdict pages on this site name the best file that fits entirely, per GPU, and point out when the native 8-bit file is worth trying.
Sources
- Ada Lovelace (RTX 40) fourth-generation Tensor Cores added the FP8 Transformer Engine first introduced in Hopper H100, so Ada has hardware FP8 tensor … www.nvidia.com →
- CUDA compute capability: RTX 50 / RTX PRO 6000 Blackwell = 12.0, RTX 40 = 8.9, RTX 30 = 8.6, RTX 20 = 7.5. developer.nvidia.com →
- ComfyUI's supports_fp8_compute() is True on NVIDIA GPUs with compute capability >= 8.9 (Ada, Blackwell) and, with PyTorch >= 2.7 and ROCm >= 6.4, on A… raw.githubusercontent.com →
- When the GPU lacks FP8 compute, ComfyUI disables the float8_e4m3fn/float8_e5m2 quantized paths and runs those layers with full-precision matmul after … raw.githubusercontent.com →
- Blackwell (GeForce RTX 50) Tensor Cores add FP4 compute. In NVIDIA's TensorRT test on an RTX 5090, FLUX.1-dev at FP4 took about 42% less time than at … developer.nvidia.com →