Tiny random Qwen3.5 model, quantized with compressed-tensors (pack-quantized int4,
group_size=16), for testing Optimum Intel's compressed-tensors export support.
Only the language-model attention/MLP linears are quantized; the vision tower, lm_head,
and the linear-attention in_proj_a/in_proj_b gating projections are left unquantized,
mirroring the ignore pattern used by real checkpoints such as cyankiwi/Qwen3.5-4B-AWQ-4bit.