This is a broken export, kept for reference only.

Asymmetric (int4_asym) weight compression on this model causes a floating-point exception (FPE) crash when running tool-call prompts on Arc B70 (OpenVINO GenAI). In short: ASYM quantization is unreliable at this compression ratio for this model/hardware combo.

Tracked upstream: https://github.com/openvinotoolkit/openvino.genai/issues/3870

Use the symmetric (int4_sym) export instead — same calibration setup, no crash: Qwen3.6-27B-r06-sym-no-mtp / Qwen3.8-27B-int4-sym in the original export directory.

OpenVINO IR export (language model, text/vision embeddings, tokenizer, detokenizer) for a Qwen3-VL 27B model, weight-compressed with mixed int4_asym/int8_asym precision (57% int4_asym / 43% int8_asym by parameter count).