ThereforeGames
afc8d7ff01
Add W6A8 checkpoint loading, conversion, inference, LoRA requantization, compile compatibility, and export, plus architecture-aware w4a8_mixed selection. Add MOSS-Image detection and quantization presets, update documentation, tests, and version to 2.4.0.
1.7 KiB
1.7 KiB
Checkpoint Sources
Pre-quantized checkpoints avoid local conversion time. On-the-fly quantization remains more flexible but requires loading the source weights and converting them locally.
Vistralis
| Model | Link |
|---|---|
| FLUX.2-klein-base-9b | Download |
| FLUX.2-klein-base-4b | Download |
| FLUX.2-klein-9b | Download |
| FLUX.2-klein-4b | Download |
Additional Checkpoints
| Model | Link |
|---|---|
| Chroma1-HD | Download |
| Z-Image-Turbo | Download |
| Anima | Download |
Grouped INT8
| Model | Format | Link |
|---|---|---|
| MOSS-Image SFT | Native ConvRot W8A8 | Download |
| MiniMax H3 FL2VA pruned | Native W6A8 | Download |
| MiniMax H3 FL2VA pruned | Legacy-metadata mixed W4A8 | Download |
The Kijai w4a8_mixed file uses legacy _quantization_metadata and stores its transformer blocks as Comfy's native asym_w4a8_int8 format. It is not a separate kernel format.