license: other
license_name: minimax-h3-community-license-agreement
license_link: LICENSE
base_model: MiniMaxAI/MiniMax-H3
base_model_relation: quantized
library_name: gguf
pipeline_tag: image-text-to-video
tags:

• sakura
• gguf
• stable-diffusion.cpp
• minimax-h3
• text-encoder
• uncensored
• abliterated
• not-for-all-audiences

Sakura — MiniMax-H3 text encoder (Qwen3-VL-32B, 50 layers), uncensored (GGUF)

For adults only; you are responsible for what you generate and for the laws that apply to you. The language weights come from a model whose refusals were removed; what this changes in the video was not measured. Do not use it to depict minors in any sexual or abusive context, or to create sexual or defamatory images of real persons without their consent. Do not put it into a public service without your own moderation.

The text encoder of MiniMax-H3 (the first 50 language layers of Qwen3-VL-32B plus its vision tower, passed to the video model with --llm) with the language weights of an abliterated (uncensored) Qwen3-VL-32B-Instruct, as GGUF for stable-diffusion.cpp. Independent work, not made, reviewed or endorsed by MiniMax, Comfy-Org, Qwen / Alibaba or llmfan46.

Where the weights come from

• The text encoder shipped inside MiniMax-H3 is the unmodified Qwen3-VL-32B-Instruct: the sha256 of all 14 text_encoder shards of MiniMaxAI/MiniMax-H3 equal those of Qwen/Qwen3-VL-32B-Instruct. MiniMax-H3 uses it truncated to 50 language layers and without the final language-model norm.
• We built the encoder from an abliterated checkpoint of that model made by a community member (llmfan46's abliterated Qwen3-VL-32B-Instruct, Apache-2.0, revision c44b949b30d111666a5ed9851c5cd633ed39b070; sha256 of both shards checked against the Hub). We did not run the ablation ourselves for this repository.
• Tensors were copied one by one into the layout of the Comfy-Org H3 encoder file (902 tensors; names, dtypes and shapes equal that file; the 14 layers 50 to 63, the final norm and the output head are dropped), then converted with sd-cli -M convert.

What you need to run it

sd-cli -M vid_gen --diffusion-model <denoiser>.gguf --vae minimax_h3_video_vae_fp16.safetensors --audio-vae minimax_h3_audio_vae_fp32.safetensors --llm Sakura-MiniMax-H3-TextEncoder-Uncensored-Q4_K_M-14.05GiB.gguf -p "A red fox walks through fresh snow in a quiet pine forest at sunrise, soft golden light, cinematic tracking shot, realistic motion, gentle wind in the trees." --cfg-scale 1.0 --steps 8 -W 480 -H 288 --video-frames 22 --fps 24 --diffusion-fa --rng cpu -s 42 -o frames/f_%03d.png

Memory: on our machine (32 GiB graphics memory plus 31.6 GiB system memory) the Q8_0 denoiser together with the Q8_0 encoder did not fit and failed with ErrorOutOfDeviceMemory; a Q4_K_M encoder of the same size as the file here (built without importance matrix) ran with every denoiser we built (Q4_K_M and Q5_K_M with and without protected tensors, Q6_K and Q8_0; 60 to 94 seconds for 22 frames at 480x288, wall-clock including model loading). This Q4_K_M file ran once, with an unprotected Q4_K_M denoiser (57 seconds). The Q8_0, Q6_K and Q5_K_M files of this repository each ran once with the published Q4_K_M denoiser of Sakura-MiniMax-H3-GGUF and produced a coherent clip (233, 91 and 75 seconds).

Example clip

One prompt, seed 42, 22 frames at 480x288 (less than one second at 24 fps), 8 steps, this Q4_K_M encoder with an unprotected Q4_K_M build of the denoiser (made by us like the Q4_K_M file of Sakura-MiniMax-H3-GGUF but without its protected tensors; not published). With the published Q4_K_M denoiser the frames differ. Decoded frames only, no audio; a tiny test render, not a statement of the full quality.

What "uncensored" means here, and what we could not measure

The language weights come from llmfan46's abliterated Qwen3-VL-32B-Instruct (linked above). Its author reports 4/100 refusals against 99/100 for the original model and a KL divergence of 0.0421 (their measurement with Heretic; not reproduced by us).

Our own check is not a valid measurement of the abliterated model. We ran the 104 held-out harmful prompts of the Heretic harmful_behaviors test split (plain chat, no system prompt, temperature 0, 120 tokens, llama.cpp b11524) on the full 64-layer model as Q8_0. The unmodified Qwen3-VL-32B-Instruct (Q8_0, unsloth) refused 104 of 104. Our own Q8_0 build of the abliterated checkpoint (converted by us with llama.cpp's convert_hf_to_gguf, without the model's chat template) gave mostly incoherent text: 68 answers start with repeated Chinese tokens, 11 more contain digit strings of 15 or more characters, 2 more repeat one phrase, and only 23 are readable. Its refusal count of 1/104 therefore does not show that the model answers. We did not find out whether this comes from our conversion or from the checkpoint. The encoder files here were built directly from the checkpoint's safetensors, not from that GGUF, and the clips made with them are coherent.

The video model reads hidden states of layer 50 of the encoder, not generated text, and we did not compare videos made with this encoder against videos made with the unmodified encoder. "Uncensored" describes where the weights come from, not a measured change of the video model's behaviour.

Quantization

• Importance matrix: own, computed with llama-imatrix on the full model (Q8_0) over 50 chunks of 512 tokens of calib_mix_v1.txt (English / German / code mix), tensor names mapped to the layout of the encoder file for layers 0 to 49. (The matrix published by MarxistLeninist contains encoder entries with a call count of 1 (the denoiser entries: 4), and under a name prefix that sd-cli -M convert -m does not apply; we did not use it for the encoder.) The file made with it differs from the plain Q4_K_M (different sha256); we did not measure the quality gain. The matrix was computed on the same Q8_0 build whose chat answers are mostly incoherent (see above); we did not check whether its statistics are representative.
• Effect on the video (one prompt, seed 42, 22 frames at 480x288, 8 steps; PSNR / SSIM of the decoded frames, mean over 22 frames):
  • with an unprotected Q4_K_M denoiser (built by us, not published): this Q4_K_M encoder against a Q4_K_M encoder without importance matrix: 20.5 dB / 0.70; in the same setting, the Q4_K_M encoder without importance matrix against Q5_K_M and Q6_K encoders without importance matrix: 21.4 dB / 0.83 and 20.0 dB / 0.67;
  • with the published Q4_K_M denoiser of Sakura-MiniMax-H3-GGUF: the Q8_0, Q6_K and Q5_K_M files of this repository against a Q4_K_M encoder without importance matrix: 23.4 dB / 0.87, 19.5 dB / 0.75 and 21.1 dB / 0.84.
  • with the published Q4_K_M denoiser, against our Q8_0 encoder as reference: the Q5_K_M file 24.3 dB / 0.89, a Q4_K_M encoder without importance matrix 23.4 dB / 0.87, the Q6_K file 22.0 dB / 0.83.
    The encoder quantization visibly changes the result, and the order is not simply by bit width: with one prompt and one seed, the Q5_K_M file came closest to the Q8_0 encoder and the Q6_K file was the farthest of the three; we do not read more into this than that. This Q4_K_M file was not run with the published Q4_K_M denoiser.

What we changed (license section III.2)

The .gguf files are modified files: language weights of an abliterated Qwen3-VL-32B-Instruct, first 50 layers, vision tower taken from the same checkpoint (its model card lists only attn.o_proj as modified), converted from safetensors to GGUF and quantized as listed.

License and use

MiniMax H3 Community License Agreement (LICENSE) including its Acceptable Use Policy (Exhibit A). By downloading or using these files you agree to LICENSE, including Section V and Exhibit A; these restrictions apply to you (among others: no use outside the Applicable Territory, no use to circumvent safeguards, no exploitation or harm of minors, disclosure of machine-generated content in public environments, no impersonation without consent). Section V.3 forbids using the models or their outputs to improve any other AI model; Section V.5 requires safeguards if you offer generation as a service; commercial use above 20 million US dollars of yearly revenue needs a separate authorization from MiniMax, and commercial products must display "MiniMax H3" (Section IV). The language weights are derived from Qwen3-VL-32B-Instruct through the abliterated checkpoint, both Apache-2.0 (LICENSE-Apache-2.0); we distribute these files as MiniMax H3 Works under the MiniMax license as well. Pass LICENSE, LICENSE-Apache-2.0 and NOTICE on with any redistribution.

Credits

• Model: MiniMaxAI/MiniMax-H3 (MiniMax); encoder layout reference: Comfy-Org/MiniMax-H3; language model: Qwen/Qwen3-VL-32B-Instruct (Qwen / Alibaba).
• Abliterated checkpoint: llmfan46.
• Conversion and inference: leejet/stable-diffusion.cpp (MIT); imatrix: ggml-org/llama.cpp (MIT).

sd-cli -M vid_gen --diffusion-model <denoiser>.gguf --vae minimax_h3_video_vae_fp16.safetensors --audio-vae minimax_h3_audio_vae_fp32.safetensors --llm Sakura-MiniMax-H3-TextEncoder-Uncensored-Q4_K_M-14.05GiB.gguf -p "A red fox walks through fresh snow in a quiet pine forest at sunrise, soft golden light, cinematic tracking shot, realistic motion, gentle wind in the trees." --cfg-scale 1.0 --steps 8 -W 480 -H 288 --video-frames 22 --fps 24 --diffusion-fa --rng cpu -s 42 -o frames/f_%03d.png