F2LLM-v2: Eight-Size Multilingual Embedding Family (80M–14B) for General-Purpose RAG

arXiv 2603.19223·medium signal

F2LLM-v2 releases 8 general-purpose multilingual embedding models spanning 80M to 14B parameters, trained on a newly curated high-quality multilingual corpus. The full size spectrum lets builders select the throughput/quality tradeoff their stack requires without switching providers. Directly actionable for any team running multilingual retrieval or semantic search pipelines.