BidirLM Omni 2.5B

BidirLM-Omni-2.5B-Embedding is a 2.5B parameter bidirectional encoder that jointly embeds text, images, and audio into a shared 2048-dimensional representation space, enabling cross-modal retrieval, semantic similarity, clustering, and classification across all three modalities.

The model supports over 119 languages (inherited from the Qwen3 base and reinforced through contrastive training with 87 languages) and accepts a 32k token context for text. Images of any size and aspect ratio are resized internally; audio at any sample rate is resampled to 16 kHz. All modalities produce embeddings directly comparable via cosine similarity.

Key architectural details: the model uses mean pooling across all modalities, requires trust_remote_code=True due to its custom bidirectional omnimodal architecture, and should be run with cuDNN > 9.20.0 to avoid a known Conv3D performance regression on H100 GPUs.

Benchmark performance is illustrated below across MTEB Multilingual V2, MIEB (lite), and MAEB (beta):

The model is based on the BidirLM architecture described in arXiv:2604.02045, which transforms causal LLMs into bidirectional encoders through a combination of prior masking, linear weight merging, and multi-domain data mixture training.