NeoMME: an efficient Multimodal-native and Multilingual Encoder

Artificial intelligence research lab H Company has released NeoMME, an open-weights multimodal and multilingual embedding model that establishes a new benchmark for cross-lingual vision-language retrieval.

The release addresses a longstanding bottleneck for developers building global AI applications: traditional image-text encoders like OpenAI CLIP and Google SigLIP perform well in English, but degrade significantly when indexing visual content in French, Hindi, Japanese, or Arabic.

According to H Company's model evaluation:

"NeoMME was trained from the ground up on a balanced 32-language visual dataset. It surpasses OpenAI CLIP-Large by 14% on cross-lingual retrieval benchmarks while requiring 40% fewer parameters, making high-performance visual search accessible on commodity hardware."

For developers building search engines, recommendation systems, and multimodal RAG pipelines, NeoMME offers several immediate architectural advantages:

1. 1.Unified Semantic Space: Text in any supported language and images map directly to the same vector coordinate. An image uploaded with French tags can be queried directly in English or Hindi without intermediate translation steps.
2. 2.Postgres pgvector Compatibility: Because NeoMME outputs standard 768-dimensional normalized vectors, developers can store and query image-text embeddings directly inside Supabase PostgreSQL using HNSW indexes.
3. 3.Edge Feasibility: The quantized model weights require less than 1.8 GB of VRAM, allowing local inference on laptops and edge servers.

The model weights and inference code have been published openly on Hugging Face under an Apache 2.0 license, allowing unrestricted commercial use.

Developers on social platforms praised the model's efficiency, with engineers noting that open-weights embedding models have now caught up with proprietary closed APIs for multilingual retrieval.

Model weights, benchmark code, and interactive demos are available on Hugging Face.