FlyCockpit/GLM-5.2-NVFP4-AQLM-Abliterated-Vision

GLM-5.2 NVFP4+AQLM Abliterated Vision

Drop-in abliterated + vision weights for the MiaAI-Lab 3× DGX Spark recipe.

Construction

Piece	Source
Text + vision hybrid base	jarrelscy/GLM-5.2-NVFP4-AQLM-hybrid @ 53e0082eedebd806b63e19779c47905937d768ca
Abliteration donor	jacklarmer/GLM-5.2-Abliterated-MXFP8-NVFP4-NF3-Hybrid
Method	Tensor transplant of BF16 self_attn.o_proj.weight for layers 13–77 (65 tensors; 62 differed from base)
Vision tower + projector	Byte-identical to the jarrelscy / Baseten vision hybrid (untouched)

No re-quantization: o_proj is BF16 [6144, 16384] on both base and donor.

Intended serve stack

• Hardware: 3× NVIDIA DGX Spark (GB10 / sm_121 / aarch64)
• Image: ghcr.io/miaai-lab/glm-5.2-nvfp4-triple-dgx-sparks:k12l1-vision
• Launcher: MiaAI-Lab ./start.sh with MM_ENCODER_TP_MODE=data, TP=3
• Playbook: https://github.com/FlyCockpit/GLM-5.2-Abliterated-Vision-3x-DGX-Sparks

Warning

This model has reduced refusal behavior relative to the stock hybrid. You are responsible for lawful use and your own evaluation.

License

Packaging Apache-2.0. Base hybrid, vision components, and upstream GLM-5.2 weights retain their original licenses/terms (see jarrelscy / DeepSeek-adjacent upstream cards). Apache-2.0 is used for this packaging where compatible with those terms.

Credits

• MiaAI-Lab triple-Spark recipe
• jarrelscy NVFP4+AQLM hybrid + vision graft
• jacklarmer abliterated donor
• Baseten / community MoonViT vision path