Why this matters Most full Qwen3‑VL‑32B checkpoints require large VRAM budgets and heavy BF16 shards. This package splits the model into a ComfyUI conditioning encoder (layers 0–49, vision tower retained in BF16) and an optional generation tail (layers 50–63) quantized with learned row-wise INT8 ConvRot. That trade lets users run multimodal conditioning and prompt enhancement workflows on ~32GB GPUs while keeping the vision tower exact.
What Sets It Apart
- Split packaging: a 24.55 GiB INT8 conditioning checkpoint and a separate 7.09 GiB INT8 generation tail let ComfyUI load only what is needed for encode vs. generation, avoiding duplication of embeddings or the vision tower.
- Quantization approach: learned ConvRot row-wise INT8 matrices for most transformer weights, a simple INT8 token embedding layout for ComfyUI compatibility, and selective BF16 retention of 551 tensors (including the full vision tower and all norms) to preserve quality where it matters.
- ComfyUI integration: designed to be selected via CLIPLoader type
minimaxand used with the MiniMax‑H3 Prompt Enhancer node; the tail is loaded temporarily for generation and unloaded afterward to leave the conditioning CLIP unchanged. - Runtime-validated: tested on a 32 GB RTX 5090 with PyTorch/CUDA combinations, showing finite conditioning outputs, measured VRAM usage, and correct layer counts after tail unload.
- Provenance and tradeoffs: derived from an "uncensored/Heretic" ARA-edited upstream source that reduces refusal rates; ablation edits are present but may alter model quality and safety behavior.
Who it's for — and tradeoffs
Great fit if you need to run Qwen3‑VL multimodal conditioning or prompt enhancement inside ComfyUI on a single high‑memory consumer GPU (≈32 GB) and want a smaller INT8 footprint while preserving the vision tower exactly. It is also appropriate when you want an optional generation tail that avoids duplicating large BF16 shards. Look elsewhere if you require an out‑of‑the‑box Hugging Face Transformers generation repo, a full BF16 copy of the model, or strict safety/refusal behavior guarantees — this package is a ComfyUI checkpoint bundle derived from an uncensored finetune and explicitly not a complete generation repository. Also note recommended ComfyUI/comfy‑kitchen and CUDA/PyTorch compatibility for best performance.
