Models.md
June 8, 2026 · View on GitHub
Pre-quantized checkpoints were recommended for most architectures, but on-the-fly quantization with ConvRot is better in all cases. However, ConvRot is also a little slower, so these prequantized models are still useful. Avoid using INT8 Tensorwise models.
Shoutout to vistralis for these:
| Model | Link |
|---|---|
| FLUX.2-klein-base-9b | Download |
| FLUX.2-klein-base-4b | Download |
| FLUX.2-klein-9b | Download |
| FLUX.2-klein-4b | Download |
ConvRot:
| Model | Link |
|---|---|
| Ideogram-4 | Download |
| LTX2.3 10Eros | Download |
| Sulphur2 Base (LTX2.3 Finetune) | Download |
| Chroma1 HD | Download |
| Ernie Image | Download |
| Anima Preview 3 | Download |
| Flux 2 Klein Base | Download |
| LTX2.3 Dev | Download |
| LTX2.3 Distilled | Download |
| WAN 2.2 | Download |
Outdated int8 models:
| Model | Link |
|---|---|
| Chroma1-HD² | |
| Z-Image-Base¹ | |
| Z-Image-Turbo² | |
| Anima | Download |
¹Z-Image Base weights have been Deprecated in favor of Convrot OTF, which is higher quality.
²Tensorwise models are worse than on the fly quantization since we switched to row-wise INT8