bridgequant-libero-7m

BridgeQuant 7M (Theia-tiny + pruned BERT-tiny, hybrid attention/short-conv trunk, short-conv flow expert) trained from scratch on the four LIBERO suites, no distillation. LIBERO-500 mean-of-3 0.948 (the distilled version bridgequant-libero-7m-kd reaches 0.967).

Files: model.safetensors, norm_stats.json, bridgequant_vla.yaml / config.json / metadata.json (the bundle load_trainable_weights expects), train_config.yaml (the experiment YAML; evaluation and deployment use the same YAML).

Code: https://github.com/VLAOpt/bridgequant-async