Look at a 2026 Comfy-Org repo — Krea 2, Qwen-Image 2.1, MiniMax H3 — and next to the familiar BF16 and FP8 files you now find int8_convrot, and in some repos nvfp4 and mxfp8. Krea 2 has all of them. Here is what they mean in practice.
INT8
An 8-bit format, about the same size as FP8. Krea 2 Turbo: FP8 13.1 GB, INT8 13.5 GB. For some new models (Qwen-Image 2.1, Z-Image) it is the only 8-bit file Comfy-Org publishes. ComfyUI computes INT8 natively on NVIDIA and AMD GPUs — including RTX 20 and RTX 30, which have no FP8 — but not on Intel. Reports on it are mixed: one community comparison ranks its quality above FP8, while an open ComfyUI issue reports it much slower than FP8 on a data-centre GPU. On this site INT8 counts as 8-bit quality and is used when it is the only 8-bit file.
NVFP4
A 4-bit floating-point format from NVIDIA. Real files are about 30–35% of the 16-bit size (4-bit values plus scales, with some layers kept at higher precision): Krea 2 Turbo NVFP4 is 26.3 → 7.7 GB. Its point is speed on RTX 50 GPUs, which compute FP4 natively: in NVIDIA's TensorRT test on an RTX 5090, FLUX.1 [dev] at FP4 took about 42% less time than at FP8 (3.9 s vs 6.7 s for 30 steps) — roughly 1.7 times the speed. ComfyUI only switches FP4 compute on for Blackwell-generation GPUs. On older GPUs such a file may still load, but you lose the reason to use it.
MXFP8
Another 8-bit variant: a "microscaling" format with a shared scale for every block of 32 weights. Same size class as FP8. ComfyUI computes it natively only on RTX 50 (Blackwell) with a recent PyTorch; on RTX 40 it is emulated, so take the plain FP8 file there.
Which should you download?
- RTX 50: FP8 if it fits; NVFP4 if you want speed and can accept 4-bit quality.
- RTX 40 and RX 9000: FP8 if it fits; on NVIDIA with current ComfyUI, even if it only nearly fits.
- RTX 30 / RTX 20: Q8_0 GGUF, or the INT8 file, which ComfyUI can compute natively on these GPUs.
- Everything else: Q8_0 GGUF if it fits, otherwise the biggest GGUF that does.
Sources
- 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 →