Free Embeddings and Reranking APIs in 2026

Every retrieval-augmented generation (RAG) app needs two cheap, high-volume calls: turning text into vectors (embeddings) and, often, reordering search hits (reranking). Five verified offers in the catalog cover both for free. Two give a one-time token balance (Voyage AI and Jina), three give rate-limited free access with no published token count (Cohere, Mistral and Gemini). Numbers below come from the catalog rows, verified on 2026-10-12 and checked for this guide on 2026-10-18; where the catalog has no figure, the table says "not recorded".

Compare the free embedding APIs

Vector dimensions are not recorded in the catalog for any of these models; check the provider's model page before you size your index.

Published rate limits

• Voyage AI: without a payment method, free trial accounts get 3 requests and 10,000 tokens a minute, which makes the 200M tokens slow to use. Adding a card moves you to Usage Tier 1 (the free tokens still apply): 2,000 requests a minute for every model listed; 3M tokens a minute on voyage-4-large, 8M on voyage-4 and voyage-code-4, 16M on voyage-4-lite, 2M on rerank-3 and voyage-multimodal-3.5 (checked 2026-10-21).
• Jina AI: the offer's limits say a free key gets 100 requests and 100,000 tokens a minute on Embeddings and Reranker (Reader: 500 requests a minute), plus an IP cap of 10,000 requests per 60 seconds.
• Cohere: trial keys are rate limited; the catalog records 2,000 embed inputs a minute and 10 rerank requests a minute.
• Mistral and Gemini: per-model limits apply; the catalog does not record embedding-specific numbers.

Which one to pick

• Most free volume: Voyage AI. 200M tokens per model is enough to embed a large document set several times over, and the reranker gets its own 200M. At 3 requests a minute without a card, a bulk import needs a payment method on file.
• Embeddings plus page-to-Markdown in one wallet: Jina AI, if your project is non-commercial: its free-key models are CC-BY-NC 4.0.
• Web search for the same agent: Jina's tokens also pay for its s.jina.ai search; the free web search API guide compares it with Exa, Firecrawl and Tavily.
• No balance to run out: Cohere, Mistral or Gemini, which stay free under rate limits. Cohere's trial keys may not be used for production or commercial purposes.
• Embeddings and chat on one key: OpenRouter had three free embedding models and one free reranker on 2026-10-21 (nvidia/nemotron-3-embed-1b:free, nvidia/llama-nemotron-embed-vl-1b-v2:free, liquid/lfm-2.5-embedding-350m:free, nvidia/llama-nemotron-rerank-vl-1b-v2:free). They share OpenRouter's free cap of 50 requests a day (1,000 after buying 10 credits), so batch many texts per request; the OpenRouter free models guide has the call and the limits.

Steps

1. Pick an offer from the table and create an account; none of the five asks for a card.
2. Create an API key in the provider's console (Jina generates one automatically on its product pages) and store it in an environment variable.
3. Embed a small sample first and check the vector length the model returns; size your vector index to match.
4. Embed documents in batches and stay under the per-minute limits above. Re-embedding everything after a model change costs the full token count again, so pick the model before you index a large corpus.
5. If you rerank, send only the top vector hits to the reranker, not the whole index: reranking spends tokens on every document you send.

When to self-host

When you outgrow a free balance or need commercial rights, run an open embedding model yourself. The catalog links Text Embeddings Inference (Apache-2.0) as the self-hosted stand-in for Voyage AI and Jina; it serves open embedding and reranker models on a CPU or an NVIDIA GPU with CUDA compute capability 7.5 or newer. For Cohere, Mistral and Gemini the catalog points to Ollama (MIT), which runs open chat and embedding models locally. Voyage's own models are not open, so self-hosting means switching models and re-embedding.

Pair it with a free vector database

The free vector database guide compares seven free tiers (Qdrant, Zilliz, Weaviate, Upstash Vector, Cloudflare Vectorize, Supabase pgvector, Turso) by how many vectors they hold. The RAG stack builder picks a free embedding model, a vector store, an LLM with a large context window and a host from the catalog in one go, using only listings that are public. For the LLM side, see free LLM APIs.

Limits and the catch

• One-time balances do not refill. Voyage AI's and Jina's free tokens are spent once; older Voyage generations (voyage-3.x, voyage-context-3, voyage-multimodal-3, rerank-2.x) get no free tokens at all.
• Non-commercial terms. Jina's free-key models are CC-BY-NC 4.0 and Cohere's trial keys are not permitted for production or commercial use.
• Free tiers can change. Rate-limited tiers are set per model and can change without notice; offers are re-checked monthly.