Voyage Code 4 is Voyage AI's code-retrieval embedding model, trained so natural-language queries such as bug reports find the right code for coding agents. It succeeds voyage-code-3 at a third lower price.
Pick it if
you index source code for coding agents or natural-language code search
you want Voyage's reported 27.54% gain over voyage-code-3 on agentic retrieval
you need 256-2048 Matryoshka dimensions to trade index size for accuracy
Skip it if
your corpus is general prose; voyage-4-lite costs a sixth as much
you need open weights to embed code on your own hardware
Jina v5 Text Small is Jina AI's 677M-parameter multilingual text embedding model, built on Qwen3-0.6B-Base and distilled from Qwen3-Embedding-4B. It is the larger of the two v5 text models, above v5-text-nano; v5-omni-small is the multimodal counterpart.
Pick it if
you embed long documents, up to 32K tokens per input
you need multilingual retrieval with vectors truncatable from 1024 down to 32
you want open weights to self-host for non-commercial work
Skip it if
you need image input; v5-omni-small covers multimodal at the same price
short inputs suffice and v5-text-nano costs less per token
FAQ
Which is cheaper, Voyage Code 4 or Jina v5 Text Small?
Jina v5 Text Small is cheaper: $0.05/M against $0.12/M.
Which handles longer documents?
Both take up to 32k tokens per text.
Can I switch from Voyage Code 4 to Jina v5 Text Small without re-indexing?
No. Vectors from different embedding models are not comparable, so a switch means re-embedding the whole corpus.
Try either model
Switch between models inside the same Octomind session.