Best Free LLM APIs for Embeddings

Embedding models convert text into fixed-size vectors. Unlike chat models, they don't generate text — they produce numerical representations for comparison and retrieval:

• Embedding dimension — The vector size. Common dimensions: 384 (lightweight), 768 (balanced), 1024–4096 (high precision). Higher dimensions capture more nuance but require more storage and slower similarity search. For most RAG applications, 768–1024 dimensions work well.
• Max input tokens — The longest text chunk the model can embed in one call. 512 tokens handles paragraphs; 8192 tokens handles entire documents. For long documents, you can either chunk them or use a model with a high token limit.
• Normalization — Most embedding models L2-normalize their outputs, making cosine similarity equivalent to dot product. This matters for vector database setup — check if you need to normalize post-hoc.
• Multilingual support — Some embedding models only work well for English. If you need multilingual semantic search, look for models with explicit multilingual training (e.g., Cohere embed-multilingual, multilingual-e5).
• Task-specific embeddings — Some models are optimized for specific tasks: classification, clustering, retrieval, or semantic textual similarity (STS). General-purpose embedders work for most use cases, but check the model card if you have a specialized need.

Embedding model selection depends on your retrieval workflow:

• Building a RAG system? → Prioritize retrieval quality. Cohere Embed (via free trial) and NVIDIA NV-Embed are strong choices. Use 768+ dimensions for good recall.
• Semantic search over documents? → Max input tokens matters — you want to embed large chunks without splitting mid-paragraph. NVIDIA NIM embedding models typically support 512–2048 token inputs.
• Multilingual search? → Cohere embed-multilingual-v3 covers 100+ languages. Check if your languages are supported before committing.
• Lightweight / edge deployment? → Cloudflare Workers AI offers embedding models (bge-base-en-v1.5, 768 dims) with 10K free requests/day. Good for smaller projects.
• Classification or clustering? → General-purpose embedders work well. You can always fine-tune on your labeled data later.

Note: embedding APIs have different authentication and SDK setups than chat APIs. NVIDIA NIM uses the same API key for both. Cloudflare Workers AI uses a different endpoint pattern. Always check the model detail page for exact base URL and code examples.