Nomic Embed

Full reproducibility set Nomic Embed apart at release: it was the first text embedding model to ship open weights, open training data, and open training code while beating OpenAI's text-embedding-ada-002 on MTEB. The family covers v1 and v1.5, both 8,192-token long-context models of roughly 100M parameters, with v1.5 adding Matryoshka Representation Learning so embeddings can be truncated from 768 down to 64 dimensions with graceful degradation, plus the mixture-of-experts multilingual v2 and matching vision models for multimodal retrieval. The contrastors repository holds the training stack, including masked language model pretraining, CLIP-style contrastive learning, GradCache for large batch training in constrained memory, and Flash Attention, with CUDA 11.8 or newer required for training runs. Inference is light enough for CPU, and the models ship as default embedders in Ollama and GPT4All. Code and the v1.5 weights are Apache-2.0, and hosted inference is available through Nomic's paid Atlas platform for teams that do not want to self-host.