Newsroom

We brought multimodal embeddings to llama.cpp and GGUF, and uncovered a few surprising issues along the way.

Code generation LLMs → code embeddings: 0.5B/1.5B models achieve SOTA performance across 25 code retrieval benchmarks.

jina-code-embeddings is a novel code embedding model suite designed to retrieve code from natural language queries, perform technical question-answering, and identify semantically similar code snippets across programming languages. It makes innovative use of an autoregressive backbone pre-trained on both text and code, generating embeddings via last-token pooling. We outline the training recipe and demonstrate state-of-the-art performance despite the relatively small size of the models, validating this approach to code embedding model construction.

Jina MCP streamlines agent development by connecting our APIs to any LLM, reducing custom code and improving reliability of the workflow.

4000 tokens/sec for a 3B-parameter embedding model on L4 GPU is probably as fast as you'll get with llama.cpp. Or is it?

Sharing what we saw and learned at SIGIR 2025, feat. CLIP-AdaM, RE-AdaptIR and evaluations for LLM-based retrieval systems.

Image resolution is crucial for embedding visually rich documents. Too small and models miss key details; too large and they can't connect the parts.

JinaVDR is a new benchmark spanning 95 tasks across 20 languages for visual document retrieval, soon on MTEB.

While others rely on prompt tuning and hope for the best, you should learn submodular optimization that provides a principled framework with theoretical guarantees for better context engineering.

Many know the importance of query diversity in DeepResearch, but few know how to solve it rigorously via submodular optimization.