qwen_quants

Qwen Quants: INT8 (ConvRot) text encoders, incl. Heretic variants

A collection of Qwen-family text encoders / language models for use as text encoders in image and video diffusion workflows (e.g. ComfyUI). Most files are "heretic" (decensored) versions of the base models, and most are quantized to INT8 with ConvRot.

Note: Replace the TODO items below with your own details before publishing.

Files

What is "heretic"?

These models were processed with Heretic, an automated tool that removes refusal behavior from language models via directional ablation (abliteration), with parameters tuned automatically to keep the model close to the original while reducing refusals.

TODO: for each heretic model, add (if you have them) the source heretic checkpoint you started from (your own run or someone else's repo, with a link), the refusal count and KL divergence reported by Heretic, and what "v2" and "merged" mean for the Qwen3.5-9B file.

What is "int8-convrot"?

TODO: describe in one or two sentences what you did, for example the quantization method or tool used, which layers are quantized and which are kept in higher precision (embeddings, norms, etc.), and any calibration data. Link to the method or code if it's public.

Usage

TODO: confirm and adjust.

1. Download the file you want.
2. Place it in ComfyUI/models/text_encoders/ (or clip/, depending on your setup).
3. Load it with TODO (which loader node / custom node is required for the INT8 ConvRot format?).
4. Pair it with a diffusion model that uses the matching text encoder:

Requirements

• TODO: custom nodes / library versions needed to load INT8 ConvRot weights
• TODO: GPU / VRAM guidance (e.g. INT8 files need roughly their file size in VRAM, plus overhead)

Quality notes

TODO: any observations, for example how INT8 compares to the full-precision version in your tests, or prompt-adherence differences. Short and honest is fine, or delete this section.

Intended use and limitations

• Intended as text encoders for generative image/video pipelines and for research on model alignment and refusal behavior.
• Heretic models have reduced or removed safety refusals. They can produce content the original models would decline. You are responsible for how you use them and for complying with applicable laws and platform rules.
• Quantization can slightly degrade quality relative to the original weights.
• These are not intended for use as general-purpose chat assistants unless you have tested them for that.

License

The base models are released by the Qwen team under their respective licenses (Apache 2.0 for the models listed above at the time of writing; please verify on each base model's page). These derived weights are distributed under the same terms. TODO: confirm the license for Krea2-Engineer-V1.

Credits

• Qwen team for the base models
• p-e-w for Heretic
• TODO: authors of the ConvRot method/tooling you used
• TODO: authors of any upstream heretic checkpoints you built on