ProMe Puts Open Models on Trial

Field notes

The project compares each listed ProMe model with an otherwise matching base model, keeping parameter counts the same and using randomized left-right placement before revealing the model identity. Its current catalogue names Qwen3.8 27B, Gemma4 32B, and Qwen3-VL 32B, with INT4 quantization noted for the latter two entries. ProMe says its models run on project-owned GPU hardware rather than through a third-party API, and describes source checking and licensing review as part of its training-data workflow.