Embedding Models for Code

Which embedding models work for the code index, what each was trained on and how they compare. For the idea, read Code search by meaning; for the setup in the Web UI, the guide Choose a model that reads code.

Keys

[Remembrances]
UseSameModel          = false          # true = the code index uses the document model
CodeEmbeddingProvider = 'ollama'       # 'ollama', 'openai', 'anthropic' or 'openai-compatible'
CodeEmbeddingModel    = 'hf.co/ggml-org/jina-embeddings-v2-base-code-Q8_0-GGUF:Q8_0'
CodeEmbeddingBaseURL  = ''             # only for 'openai-compatible' or a remote Ollama
CodeEmbeddingAPIKey   = ''             # stored encrypted

Web UI: Settings > Remembrances > Code embeddings. With UseSameModel = true, or with the code fields empty, the code index uses DocumentEmbeddingProvider and DocumentEmbeddingModel.

After changing the model, run Re-index all (Settings > Remembrances > Code indexing). Vectors made by another model are not comparable, and vectors of a different size are skipped by the search.

Models that work in Ollama

All were pulled and tested with Ollama 0.35 in October 2026. GGUF is the file format Ollama loads; hf.co/<user>/<repo>:<tag> pulls a GGUF file directly from Hugging Face.

The name to give Ollama and to type in Code embedding model:

ollama pull hf.co/brandtcormorant/CodeRankEmbed-Q4_K_M-GGUF:Q4_K_M         # CodeRankEmbed
ollama pull hf.co/brandtcormorant/CodeRankEmbed-Q8_0-GGUF:Q8_0             # CodeRankEmbed, less compressed
ollama pull hf.co/ggml-org/jina-embeddings-v2-base-code-Q8_0-GGUF:Q8_0     # Jina Embeddings v2 Base Code
ollama pull qwen3-embedding:0.6b                                           # Qwen3 Embedding 0.6B
ollama pull nomic-embed-text                                               # Nomic Embed Text v1.5

The 30 languages of Jina v2 Base Code: Assembly, Batchfile, C, C#, C++, CMake, CSS, Dockerfile, FORTRAN, Go, Haskell, HTML, Java, JavaScript, Julia, Lua, Makefile, Markdown, PHP, Perl, PowerShell, Python, Ruby, Rust, SQL, Scala, Shell, TypeScript, TeX and Visual Basic.

Smaller files of Jina v2 Base Code (from 83 MB) are in hf.co/second-state/jina-embeddings-v2-base-code-GGUF, with tags such as Q4_K_M.

A small comparison

This is a home-made check, not a benchmark. 48 snippets (six tasks, each written in Python, Go, TypeScript, Java, Rust, PHP, C# and shell) with meaningless function names, so the model has to read the code. For each task a plain-English question is asked and we count how many of the eight right snippets are among the first eight results. Speed is 1,500-character pieces of real source files embedded per second on one desktop graphics card; read it as a ratio between models, not as what you will get.

Reading it:

• A general text model is the weak choice for code, and the gap is largest in languages it saw little of.
• CodeRankEmbed is the lightest. It is strong in the six languages it was trained on and weaker outside them.
• Jina v2 Base Code covers the most languages for its size and speed.
• Qwen3 Embedding matches it and also reads many human languages, at four times the cost in time and space.
• The more compressed CodeRankEmbed file gave the same results as the larger one.

CodeRankEmbed is designed to receive questions with the prefix Represent this query for searching relevant code: . Pando does not add it. In the check above the prefix raised its result from 88 % to 92 %.

Models that did not work or were not tried

Related

• Remembrances configuration: the rest of the index and search keys.