I build, break, and deploy AI systems.

Remove refusal behavior from a 30.7B multimodal model without losing a deployable format set.

Applied weight-level abliteration, then carried the same model lineage through BF16, ModelOpt NVFP4, and a GGUF K-quant ladder. For multimodal NVFP4 serving, the vision tower, embedder, and lm_head stayed in BF16.

BF16 evaluation: 0/686 effective refusals across four datasets. The 20 GB NVFP4 build, reduced from 59 GB, was validated in vLLM and SGLang. GGUF: 12–31 GB with 0–1/100 hard refusals across quant levels.

Fit a 34.7B Qwen3.5-MoE reasoning checkpoint onto a single Blackwell workstation without quantizing every subsystem.

Quantized routed experts to NVFP4 with NVIDIA ModelOpt; kept attention QKV, shared experts, and the vision encoder at higher precision.

About 23 GB versus 69 GB BF16, validated to load and generate with vLLM on an RTX PRO 6000 Blackwell.

Study refusal behavior in a 428B multimodal MoE while preserving its reasoning and routing structure.

Modified attention o_proj and every residual-writing down_proj across all 60 layers; released the result as 796 GB BF16 in 59 shards.

0/16 hard refusals on the published sample; multimodal behavior, reasoning, and MoE routing were smoke-tested, not fully benchmarked.