MiniMax-H3 Text Encoder — Qwen3-VL-32B (abliterated) · NVFP4-AWQ

Quantized text encoder for the MiniMax-H3 video model, built to be served by vLLM-Omni.

Base: abliterated Qwen3-VL-32B-Instruct
Layers: 64 → 50 (H3 only consumes the hidden state after layer 50; lm_head and the final norm are dropped)
Quantization: NVFP4 (W4A16, group_size 16) via NVIDIA ModelOpt, with AWQ pre_quant_scale on down_proj / o_proj (per-layer α searched on real activations; median down_proj≈0.4, o_proj≈0.5). embed_tokens → FP8 (e4m3). Vision tower and all norms kept BF16.
Format: ModelOpt convention — vLLM-Omni's W4A16_NVFP4 path requires pre_quant_scale, so this build ships it (a plain NVFP4 build without it will not load).
Size: ~15.7 GB.

Notes

This is an abliterated / uncensored derivative, released as a component for the MiniMax-H3 text-to-video / image-to-video pipeline.

license: apache-2.0 base_model:

Qwen/Qwen3-VL-32B-Instruct pipeline_tag: image-text-to-text library_name: transformers tags:
nvfp4
awq
modelopt
abliterated
uncensored
minimax-h3
text-encoder
vllm-omni

Quantized text encoder for the MiniMax-H3 video model, built to be served by vLLM-Omni.

Base: abliterated Qwen3-VL-32B-Instruct
Layers: 64 → 50 (H3 only consumes the hidden state after layer 50; lm_head and the final norm are dropped)
Quantization: NVFP4 (W4A16, group_size 16) via NVIDIA ModelOpt, with AWQ pre_quant_scale on down_proj / o_proj (per-layer α searched on real activations; median down_proj≈0.4, o_proj≈0.5). embed_tokens → FP8 (e4m3). Vision tower and all norms kept BF16.
Format: ModelOpt convention — vLLM-Omni's W4A16_NVFP4 path requires pre_quant_scale, so this build ships it (a plain NVFP4 build without it will not load).
Size: ~15.7 GB.

This is an abliterated / uncensored derivative, released as a component for the MiniMax-H3 text-to-video / image-to-video pipeline.