📌 Heads-up: I've been focused on still-image work lately, so I don't have the bandwidth to help with video setups like MiniMax-H3 for now. The files stay up as they are — thanks for understanding! 最近重心放在平面攝影，影片類（像 MiniMax-H3）暫時比較沒辦法協助排錯。檔案會照舊留著，謝謝體諒！
MiniMax-H3 Text Encoder — Qwen3-VL-32B (abliterated) · GGUF
GGUF (llama.cpp) builds of the MiniMax-H3 text encoder — abliterated Qwen3-VL-32B, 50 layers, Q3_K_S (≈3.53 BPW).
Files
| File | Size | Visual tower | Use for |
|---|---|---|---|
MiniMax-H3-TextEncoder-Qwen3VL-32B-abliterated-Q3_K_S.gguf |
~11.5 GB | ✗ (text only) | plain llama.cpp text-encoder use |
MiniMax-H3-TextEncoder-Qwen3VL-32B-abliterated-Q3_K_S_vis.gguf |
~12.6 GB | ✓ merged in | ComfyUI (MiniMax-H3 pipeline) |
For ComfyUI use the _vis file. ComfyUI detects the H3 text encoder via the visual tower tensors (visual.deepstack_merger_list.* + model.layers.49.*); the plain file lacks them and won't be recognized. The _vis build has the BF16 visual tower merged back in.
- Layers: 50 (H3 consumes the hidden state after layer 50) · Arch:
qwen3vl· Base: abliterated Qwen3-VL-32B-Instruct.
For the vLLM-Omni-ready NVFP4-AWQ build, see pottokao/MiniMax-H3-TextEncoder-Qwen3VL-32B-abliterated-NVFP4-AWQ.
Abliterated / uncensored derivative, released as a component for the MiniMax-H3 text-to-video / image-to-video pipeline.