Qwen3.5 122B INT4 heretic quantization fits 63GB

Combining aggressive quantization with alignment ablation in one release is rare and technically interesting — most community quants pick one or the other.

• –Intel AutoRound's sign-gradient descent INT4 achieves 73% size reduction (234GB → 63GB) with KL divergence of only 0.0916 vs. the base model
• –Directional ablation via Heretic v1.2.0 cuts refusal rate from 99/100 to 9/100 test prompts by targeting `attn.o_proj` and `mlp.down_proj` layers specifically
• –Critical precision layers (vision encoder, shared expert projections, MoE routing gates, lm_head) are kept at FP16/BF16 — a thoughtful decision since shared experts fire on every token
• –vLLM-compatible with GPTQ Marlin format; includes DGX Spark-specific deployment notes (NCCL deadlock workarounds, unified memory utilization caps)
• –Low community traction so far (score 6, 2 comments) but technically solid enough to be useful for local LLM practitioners with high-VRAM setups