F2LLM V2 1.7B

F2LLM-v2-1.7B is a general-purpose, multilingual embedding model that converts text into dense vector representations for information retrieval, semantic search, and text classification. It is the 1.7-billion-parameter instruct variant of the F2LLM-v2 family, fine-tuned from the codefuse-ai/F2LLM-v2-1.7B-Preview base model and released under the Apache 2.0 license.

Capabilities and Training

Trained on a curated composite of 60 million high-quality samples from the codefuse-ai/F2LLM-v2 dataset, the model supports over 200 languages with a particular focus on mid- and low-resource languages. According to the research paper (arXiv:2603.19223), the training corpus covers 282 natural languages and more than 40 programming languages. The architecture is based on Qwen3 and produces output embeddings with a fixed dimension of 2048.

Optimization Techniques

The model integrates a two-stage LLM-based embedding training pipeline with Matryoshka Representation Learning (MRL), model pruning, and knowledge distillation, enabling efficient inference while maintaining competitive retrieval performance.

Benchmark Performance

While specific MTEB scores for the 1.7B variant are not published, the family’s largest model, F2LLM-v2-14B, ranks first on 11 MTEB benchmarks. The 1.7B model inherits the training methodology and is designed to deliver strong embedding quality at a compact size suitable for resource-constrained applications.

Model Family Overview

All models, training data, code, and intermediate checkpoints are publicly available. Intermediate checkpoints for this model are provided in the intermediate_checkpoints branch on Hugging Face.