The best embedding models, according to hackers
8 embedding models recommended in Hacker News comments over the past year, ranked by how many separate people recommend each one.
All embedding models
- EmbeddingGemmaOpen embedding model people praise for quality, low RAM use and its Apache 2.0 license. Critics note weaker accuracy than cloud models for speech and some tasks.27recommend
- TernlightTiny embedding model that runs in the browser for on-device semantic search without API calls. Caveat: heavy CPU and memory use, and failures on some browsers and phones.17recommend
- CLIP by OpenAIRecommended for image embeddings and similarity search; it is open source and good at object recognition. Critics say it struggles beyond objects and lacks temporal understanding.10recommend
- Gemini Embedding 2People recommend it for embedding text, images, audio, video and PDFs in one vector space, with a generous free tier. Critics say it gets expensive at scale.9recommend
- Qwen3-EmbeddingOpen-weight multilingual embedding model with 32K context and several sizes, popular for local retrieval. Caveat: some find its latency too high for their use case.8recommend
- ModernBERTFree open encoder model that people fine-tune for classification and embeddings; it learns quickly and works well as a starting point. Critics say it underperforms on some tasks.7recommend
- all-MiniLM-L6-v2A small, free, open embedding model that runs on-device and gives good semantic matching for its size. Critics say it is old and behind newer open-weight models.6recommend
- word2vecRecommended as a beautiful piece of science that dominated word embeddings for about five years, though one critic says the criticism of it is valid.5recommend