F2LLM V2 8B

F2LLM-v2-8B is a multilingual embedding model from the F2LLM-v2 family, designed for general-purpose text embedding tasks such as information retrieval, semantic search, text classification, clustering, and bitext mining. It supports over 200 languages (282 natural languages and 40 programming languages in the training corpus).

Architecture and Training

The model is trained on a curated composite of 60 million publicly available high-quality data samples. It employs a two-stage LLM-based embedding pipeline integrating matryoshka learning, model pruning, and knowledge distillation to deliver competitive performance with greater efficiency than previous LLM-based embedding models.

Benchmark Performance

Within the F2LLM-v2 family, the 14B model ranks first on 11 MTEB benchmarks. The 8B variant achieves state-of-the-art results for its size class, making it suitable for resource-constrained applications while retaining strong retrieval and similarity capabilities.

Model Variants

Prompt Format

For asymmetric retrieval and reranking, prepend the instruction to the query using the format Instruct: your_instruction\nQuery: and leave documents unprompted. For symmetric tasks (STS, clustering, bitext mining), the model supports both prompted and unprompted encoding.

Pipeline: feature-extraction. License: CC BY-NC-ND 4.0.