Ant Group Unveils F2LLM-v2: A Multilingual, Full-Scale Embedding Model

Overcoming the "English-centric" limitation in semantic representation has emerged as a key frontier in the evolution of large language models.

On March 26, the CodeFuse team from Ant Group and Shanghai Jiao Tong University officially released the F2LLM-v2 series of Embedding models. This series not only achieved leading performance in authoritative benchmarks but also provides a high-performance, efficient semantic representation solution for developers worldwide through a fully open-source approach.

Exceptional Performance: Achieving 11 SOTA Results on MTEB

In the authoritative MTEB benchmark for evaluating Embedding models, F2LLM-v2 demonstrated comprehensive strengths:

11 Top Rankings: It secured first place across 11 language and domain-specific leaderboards, including German, French, Japanese, and code retrieval.

A Formidable Challenger: Even its lightweight variants consistently outperformed well-known industry models of comparable size.

Extensive Coverage: The evaluation encompassed 430 diverse sub-tasks, such as medical Q&A and code retrieval, achieving full-scenario coverage.

Comprehensive Understanding: Proficiency in 282 Natural Languages and Over 40 Programming Languages

The power of F2LLM-v2 stems from its highly inclusive training foundation:

Multilingual Enhancement: It features strengthened support for medium- and low-resource languages (such as Nordic and Southeast Asian language families), enabling genuine global language coverage.

Programming Expertise: With deep understanding of over 40 programming languages like Python, Java, and Go, it is an ideal choice for developers building RAG (Retrieval-Augmented Generation) systems and code assistants.

High-Quality Data: Built upon 60 million meticulously cleaned public samples, it ensures both the purity and breadth of the model's knowledge base.

Extreme Efficiency: A Complete Model Family Scaling from 80M to 14B Parameters

To address needs ranging from mobile devices to cloud computing, the CodeFuse team developed a comprehensive model matrix:

Mobile-Optimized: Compact models from 80M to 330M parameters utilize "model pruning" and "knowledge distillation" techniques, enabling smooth operation on mobile platforms.

"Nested" Innovation: It supports dynamic dimension adjustment, allowing users to flexibly switch between 8 dimensions and full dimensions, optimizing the trade-off between inference speed and storage cost.

Fully Open Source: Transparency Setting a New Community Standard

Unlike many "black box" models, F2LLM-v2 is committed to a fully open-source philosophy:

Complete Release: All model weights for every size variant are available for download.

Detailed Transparency: A comprehensive technical report is published, disclosing the entire training methodology.

Full Reproducibility: All code and training checkpoints are released, empowering researchers globally to build upon this work for further development.

Conclusion: Breaking Boundaries to Explore AI's Infinite Potential

As another significant milestone in the CodeFuse Open Source Series, the release of F2LLM-v2