#MTEB

'IBM published two ModernBERT-based Granite R2 embeddings offering 8k context, compact architectures, and high retrieval throughput suitable for production RAG and search systems.'

'Google released EmbeddingGemma, a 308M-parameter on-device embedding model that tops MTEB scores for models under 500M and delivers low-latency multilingual retrieval suitable for offline RAG.'