Jina Embeddings v5: local setup and multilingual tests

Fahad Mirza demonstrates a local setup of jina-embeddings-v5-text-small on Ubuntu, then tests document retrieval and multilingual sentence similarity. He creates a virtual environment, installs PyTorch and Transformers, and downloads the model. The example uses a GPU with BFloat16; he says the model can also run on CPU, though the demonstration does not establish CPU performance.

The retrieval script encodes queries and documents with task-specific prompts and ranks them by cosine similarity. For a question about inflation, the presenter reports scores of 0.74 for a direct explanation and 0.39 for an indirectly related answer. He questions the ordering of some low-scoring results, then repeats the run and gets the same output. This is a small retrieval example rather than a broad accuracy benchmark.

The second test compares an English sentence about sunrise and sunset with translations in more than 50 languages, using 1024-dimensional vectors. The presenter reports similarities of 0.83 to 0.94 for several major languages and lower scores for Hausa, Malagasy, Sinhala and Kurdish. He attributes those gaps to less training data. He also describes a Qwen3-0.6B base and support for sequences up to 32K tokens. The walkthrough shows how to run models locally for embedding experiments, while its results remain specific to the examples shown.