They convert decoder architectures into encoders to optimize search.
Introducing LFM2.5-Embedding-350M and LFM2.5-ColBERT-350M: two multilingual retrieval models built for ultra-fast and accurate search across 11 languages.
> End-to-end retrieval latency as low as 1.5ms with our enterprise stack! 🚀
> Consistently best-in-class multilingual and cross-lingual performance across Arabic, German, English, Spanish, French, Italian, Japanese, Korean, Norwegian, Portuguese, and Swedish.
🧵
Users are excited about the new LFM2.5 models' ultra-fast multilingual retrieval and decoder-to-encoder conversions for ColBERT because they enable practical RAG testing and show strong robustness across languages and architectures.
No Digg Deeper questions have been answered for this story yet.
Both are built on LFM2.5-350M-Base with bidirectional patches, the first full-context encoders in the LFM family.
> Both run via llama.cpp GGUFs (CPUs, Laptops, edge devices) <10ms end-to-end query embedding latency.
> Both fly on GPUs with our custom runtime <2ms end-to-end query embedding latency.
(3/n)
Especially well-suited for short-context collections: product catalogs, FAQ, knowledge bases, support docs, and multilingual search use cases where speed and reliability matter.
LFM2.5-Embedding-350M and LFM2.5-ColBERT-350M are available now.
> Blog post: https://www.liquid.ai/blog/lfm2-5-retrievers > LFM2.5-Embedding-350M: https://huggingface.co/LiquidAI/LFM2.5-Embedding-350M > LFM2.5-ColBERT-350M: https://huggingface.co/LiquidAI/LFM2.5-ColBERT-350M > Docs: http://docs.liquid.ai
Turning a decoder into an encoder and training a late interaction model with this backbone... Maybe the propaganda is working, after all
Very cool to see new multilingual ColBERT models popping up! We might have something on this topic for you very soon... which might create even more and stronger multilingual ColBERT!
Introducing LFM2.5-Embedding-350M and LFM2.5-ColBERT-350M: two multilingual retrieval models built for ultra-fast and accurate search across 11 languages.
> End-to-end retrieval latency as low as 1.5ms with our enterprise stack! 🚀
> Consistently best-in-class multilingual and cross-lingual performance across Arabic, German, English, Spanish, French, Italian, Japanese, Korean, Norwegian, Portuguese, and Swedish.
🧵
Retrieval models force a tradeoff between speed and accuracy. These two sit at different points on that curve.
> LFM2.5-Embedding-350M: one vector per document. Fastest search, smallest index.
> LFM2.5-ColBERT-350M: one vector per token. Word-level query matching, higher accuracy, larger index.
(2/n)
@liquidai That's great, going to test it for RAG
@liquidai Interesting release 🤔
Also, it's very cool how ST/PyLate are robust to these different architectures, even when they are tweaked You can just plug things
@liquidai @Presidentlin
@liquidai Korean is huge
@liquidai Very impressive progress!
