pt810/Ovis-Omni-Embedding-3B-gptq-mixed-w2-w4-w8

Ovis Omni Embedding 3B — experimental GPTQ W2/W4/W8

Experimental calibration-based GPTQ-derived variant of ATH-MaaS/Ovis-Omni-Embedding-3B.

Thinker layers 0–13 use 2-bit weights, layers 14–26 use 4-bit weights, and layer 27 uses 8-bit weights. The multimodal towers and non-linear/special modules remain in the original precision. Group scales were recomputed from the original BF16 weights after the compressor export produced invalid scales.

Compatibility warning: this W2 mixed artifact is published for experimentation, but vLLM 0.30.0 rejects its W2 packed shape in the Marlin compressed-tensors loader. Do not treat it as a vLLM-servable checkpoint. Use the separately published W4/W4/W8 vLLM-tested artifact for vLLM serving.