license: apache-2.0 base_model:
- llmfan46/Qwen3-VL-32B-Instruct-ultra-uncensored-heretic
- Qwen/Qwen3-VL-32B-Instruct library_name: gguf tags:
- gguf
- minimax-h3
- text-encoder
- abliterated
- comfyui
MiniMax-H3 Text Encoder (abliterated) for 16 GB GPUs — GGUF
⚠️ Not a general-purpose LLM. These files are a text encoder for MiniMax-H3 and nothing else: cut to the 50 layers H3 uses, with no LM head — they cannot chat or generate text. They exist so H3's 32B encoder fits a 16 GB GPU in ComfyUI.
Abliterated variant of pottokao/MiniMax-H3-TextEncoder-16GB-GGUF.
The abliteration is llmfan46's
(Qwen3-VL-32B-Instruct-ultra-uncensored-heretic):
compared tensor by tensor with the official H3 encoder, it differs only in self_attn.o_proj of layers 31–39;
those nine tensors were swapped into the official encoder, everything else is identical. Same per-tensor format
allocation as the official build.
Files (ComfyUI, visual tower included)
| File | Size | Distance from the official BF16 encoder (FL2VA / Ref2VA) |
|---|---|---|
MiniMax-H3-TextEncoder-Qwen3VL-32B-abliterated-11.5G_vis.gguf | 12.65 GB | 11.1% / 11.3% |
MiniMax-H3-TextEncoder-Qwen3VL-32B-abliterated-12.5G_vis.gguf | 13.70 GB | 10.3% / 10.5% |
MiniMax-H3-TextEncoder-Qwen3VL-32B-abliterated-13.6G_vis.gguf | 14.80 GB | 9.3% / 9.5% |
Measured after the H3 DiT's own text stage (condition_proj → token_refiner, FL2VA and Ref2VA copies) on 36
held-out H3-format prompts, against the official BF16 encoder. So these numbers include the abliteration itself,
not only quantization: the official build at the same sizes scores 7.9% / 6.4% / 4.9%. Read together, the abliteration
alone moves what the DiT sees by roughly 8% — it does change the conditioning; whether that helps or hurts a given
video is not measured here.
The earlier MiniMax-H3-TextEncoder-Qwen3VL-32B-abliterated-GGUF Q3_K_S (same abliteration, every tensor Q3_K) scores
14.4% / 14.7% on the same test.
Use in ComfyUI
- ComfyUI-GGUF
CLIPLoaderGGUF, typeminimax. - Each file was loaded this way and encoded 12 held-out prompts: all tensors dequantize on the GPU (K-quants only, no CPU fallback), ~25–32 s to load, ~6 s per prompt.
- Visual tower merged as F16 (ComfyUI detects the H3 encoder by it); the 5-D
patch_embedweight is stored the ComfyUI-GGUF way, so currentggufversions read it. - 16 GB cards → use the 11.5G file. Tested on an RTX 5070 Ti 16 GB (ComfyUI, H3 FL2VA turbo 4-step DiT, euler + sigma shift 12/3, 1344×768 × 61 frames): full clips render with default settings, ~75 s per clip, GPU peak ≈ 15 GB.
- 12.5G / 13.6G on a 16 GB card: with ComfyUI's default settings the encoder runs out of memory (
torch.OutOfMemoryError, no video). Starting ComfyUI with--reserve-vram 1.5makes it keep part of the encoder in system RAM: the encoding step then passes (first prompt 12 s / 19 s, later ones ~4 s), but a full clip with that setting was not tested — the DiT may then be pushed to offload too. On 20 GB+ cards they load normally.
Format of every tensor and how the sizes were spent: see the official repo (GRIDS.md there applies here too).
License
Qwen3-VL-32B, Apache 2.0 (the MiniMax H3 license states its encoder is Qwen3-VL-32B under Apache 2.0); abliteration by llmfan46. MiniMax H3 itself (DiT / VAE) is under the MiniMax H3 Community License.
📌 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！