Flagship 14B multilingual embedding model from CodeFuse-AI; SOTA on 11/17 MTEB benchmarks.
Access model weights, configuration files, and documentation.
See which devices can run this model and at what quality level.
The flagship of Ant Group CodeFuse-AI's F2LLM-v2 family, built on Qwen3-14B-Base and trained on a fully open 60M-sample corpus spanning 282 natural and 40+ programming languages via a two-stage contrastive + instruction-tuned recipe with Matryoshka Representation Learning. Releases all weights, intermediate checkpoints, training data, and code as a reproducible open baseline; achieves SOTA on 11 of 17 evaluated MTEB benchmarks.
View on Hugging Face