Hugging Face Blog Releases NeoMME: An Efficient Multimodal-Native and Multilingual Encoder
The Hugging Face blog announces NeoMME, a family of 260M and 800M-parameter multimodal and multilingual bidirectional encoders. NeoMME processes multilingual text and raw image patches in a single bidirectional Transformer encoder, replacing separately pretrained image and text encoders. All model checkpoints and a day-zero Hugging Face Transformers implementation are released under Apache 2.0.
What happened
Event details
The Hugging Face blog announces NeoMME, a family of 260M and 800M-parameter multimodal and multilingual bidirectional encoders. NeoMME processes multilingual text and raw image patches in a single bidirectional Transformer encoder, replacing separately pretrained image and text encoders. The post reports that NeoMME-Retriever-260M reaches 0.523, the highest score among evaluated models strictly below 800M parameters. All NeoMME model checkpoints and a day-zero Hugging Face Transformers implementation are released under Apache 2.0. The blog post credits H Company for supporting the work and providing the compute used to train NeoMME.
Assessment
Why it matters
This is a first-party Hugging Face blog post announcing a new model family (NeoMME) with concrete technical details, benchmark results, and open-source release information. It is a distinct, materially important release event.
Availability
Access notes
All NeoMME model checkpoints and a day-zero Hugging Face Transformers implementation are released under Apache 2.0.