If you want an open source vector database you can run anywhere, with a managed cloud when you want one, look first at Qdrant and Weaviate. For very large collections and a wide choice of index types, Milvus self-hosted, or Zilliz Cloud if you still want someone else to run it. If your application data already lives in PostgreSQL, pgvector may remove the need for a separate vector store; the same reasoning applies to MongoDB Atlas Vector Search for MongoDB users and to Redis for teams already running Redis. If you already operate a search cluster, Elasticsearch or OpenSearch can hold the vectors next to your text indexes.
Pinecone is a proprietary, fully managed vector database. You create serverless indexes through its API, upsert vectors with metadata and query them by similarity; since September 2026 document indexes can also hold sparse vectors and BM25 full-text fields, and Pinecone can host the embedding and reranking models. There is no open source or downloadable server edition. Pinecone runs on AWS, Google Cloud and Azure, and Enterprise customers can run it in their own cloud account through Bring Your Own Cloud (BYOC), generally available since 25 August 2026. Plans run from a free Starter plan and a flat $20 Builder plan to usage-based Standard ($50 monthly minimum) and Enterprise ($500 monthly minimum) plans billed on read units, write units and storage.
Teams that look elsewhere usually give one of these reasons. The first, third and fourth are documented by Pinecone; the second is our editorial reading.
- No self-hosted edition. Outside Enterprise BYOC, the database runs only on Pinecone's platform, so on-premises, air-gapped or strict data residency requirements are hard to meet, and the data and API cannot be moved to another host without a migration.
- A second system next to the main database. If the records you embed already sit in PostgreSQL, MongoDB or a search cluster, a separate vector service means another copy of the data to keep in sync, another bill and another security review (editorial).
- Usage-based cost. On Standard and Enterprise, read units grow with query volume and result size, which is harder to forecast than fixed capacity you size yourself.
- Legacy pod indexes. Pod-based indexes are legacy: Standard and Enterprise organisations created on or after 18 August 2025 cannot create them, and SDKs on API version 2026-07 no longer create them. Teams with pod indexes face a migration either way, which is a natural point to compare options.
For head-to-head pages, see Pinecone vs Weaviate, pgvector vs Pinecone, Milvus vs Qdrant and Elasticsearch vs OpenSearch for vector search. If running a distributed open source cluster is itself the worry, see our Milvus alternatives page, which is written for that problem.
Quick picks
Apache 2.0 Rust server with filterable HNSW, server-side BM25 hybrid queries and a free-forever single-node cluster on Qdrant Cloud.
BSD-3-Clause core with BM25 hybrid search, per-tenant shards and model provider modules, self-hosted or on Weaviate Cloud.
Managed Milvus on AWS, Google Cloud and Azure, with the option to move to self-hosted Apache 2.0 Milvus later.
Vectors and HNSW indexes in the same database and transactions as your relational data, on most managed PostgreSQL services.
Apache 2.0 k-NN vector fields and hybrid queries next to existing text indexes.
How we chose
We considered Qdrant, Weaviate, Milvus, Zilliz Cloud, Chroma, LanceDB, Vespa, pgvector, MongoDB Atlas Vector Search, Redis, Elasticsearch, OpenSearch, SingleStore and FAISS. To be included, a product had to be actively developed or sold in November 2026, store vectors with metadata and serve approximate nearest neighbour queries in production, and either be open source or self-hostable, or run inside a database that teams commonly already operate. That second rule follows the two reasons above for leaving Pinecone.
FAISS was left out because it is a library, not a database: it has no persistence layer, metadata storage or replication. Chroma and LanceDB are covered on the Milvus alternatives page, where their embedded and single-node designs answer a different question (see Chroma vs LanceDB). Vespa is a strong engine but adds schemas, tensors and ranking expressions, so we mention it only in the FAQ. SingleStore has vector indexes, but it is a commercial distributed SQL database that few teams adopt only for vectors.
For each product we recorded the licence, deployment options, pricing model, hybrid search support and its main limitation. Exact prices are only given where the vendor publishes them; usage rates change by region, so check the linked pricing pages. The order is not a ranking: entries start with open source vector databases, then the managed service built on Milvus, then vector search inside existing databases and search engines.
At a glance
| Product | Type | Licence | Deployment | Hybrid search | Main trade-off |
|---|---|---|---|---|---|
| Qdrant | Vector database | Apache 2.0 | Self-hosted; Qdrant Cloud, Hybrid and Private Cloud | Yes: RRF or DBSF fusion, server-side BM25 | Standard cloud unit rates only in the calculator |
| Weaviate | Vector database | BSD-3-Clause (wl/ folder commercial) | Self-hosted; Weaviate Cloud | Yes: BM25 plus vector fusion | In-memory HNSW needs RAM planning |
| Milvus | Vector database | Apache 2.0 | Self-hosted: Lite, Standalone, Distributed | Yes: BM25 sparse plus dense | Distributed mode needs etcd, object storage and a WAL service |
| Zilliz Cloud | Managed Milvus | Commercial (engine Apache 2.0) | AWS, Google Cloud, Azure; BYOC | Yes: multi-vector, sparse plus dense | Index choice reduced to AUTOINDEX |
| pgvector | PostgreSQL extension | PostgreSQL License | Any PostgreSQL, most managed services | Not built in; combine with PostgreSQL full-text search in SQL | HNSW index limited to 2,000 dimensions for vector |
| MongoDB Atlas Vector Search | Feature of a document database | Commercial service (server SSPL) | Atlas on AWS, Azure, Google Cloud; local Atlas deployments | Yes: vector plus full-text search | Not supported on Atlas Infinite clusters in preview |
| Elasticsearch | Search engine | AGPLv3, SSPL or ELv2; paid tiers | Self-managed; Elastic Cloud Hosted or Serverless | Yes: RRF and linear retrievers | Some vector features need an Enterprise subscription |
| OpenSearch | Search engine | Apache 2.0 | Self-hosted; Amazon OpenSearch Service; other providers | Yes: hybrid query (since 2.11) | Cluster sizing and engine choice to manage |
| Redis | In-memory database | RSALv2, SSPLv1 or AGPLv3 (8.0+) | Self-hosted; Redis Cloud; Redis Software | Yes: FT.HYBRID (since 8.4.0) | Data set bounded by memory without tiering |
Qdrant
Qdrant is an Apache 2.0 vector database written in Rust. Its HNSW index is filter-aware when payload indexes (keyword, numeric, geo, datetime, text, uuid) are created first, and its Query API combines dense and sparse prefetches with RRF or DBSF fusion, including server-side BM25. Multitenancy can use tenant partitioning, user-defined shards or tiered multitenancy.
For teams leaving Pinecone, the main point is choice of where it runs: self-hosted on Docker or Kubernetes, on Qdrant Cloud (with a free-forever single-node cluster of 0.5 vCPU, 1 GB RAM and 4 GB disk), or as Hybrid Cloud and Private Cloud in your own infrastructure. The Standard tier is billed hourly on vCPU, memory and storage, which some teams find easier to forecast than per-request units (editorial). See Milvus vs Qdrant.
- Payload indexes should exist before HNSW is built; adding them later means rebuilding the index.
- The free cloud cluster has no high availability and is deleted after 4 weeks of inactivity.
- Standard tier unit rates appear only in the pricing calculator, not in the pricing page text.
- Version 1.19 removed the legacy /search, /recommend and /discover endpoints, so older clients must move to /query.
Weaviate
Weaviate is a vector database written in Go, with most of its repository under the BSD-3-Clause licence. It offers HNSW, flat, dynamic and HFresh indexes, BM25 plus vector hybrid search with relativeScoreFusion or rankedFusion, and native multi-tenancy with one shard per tenant. Model provider modules (API-based and local) and Weaviate Embeddings can create vectors for you, which is the closest open source match to Pinecone's integrated inference.
Weaviate Cloud has a Free plan (100,000 objects, 1 collection), Flex from $45 a month with no commitment, and Premium contracts from $400 a month, billed per million vector dimensions, storage and backup; the price book changed for usage from 1 September 2026. See Pinecone vs Weaviate.
- The default HNSW index is held in memory, so large self-hosted deployments need RAM planning or compression.
- Files in the repository's wl/ folder are under a commercial licence, so check which components you use.
- Cloud billing by vector dimensions varies with index type, compression and region.
Milvus
Milvus is an Apache 2.0 vector database hosted by the LF AI & Data Foundation, with the widest documented set of index types in this list: FLAT, the IVF family, HNSW with quantization variants, DISKANN, SCANN and GPU indexes such as CAGRA. It supports BM25 full-text search as sparse vectors, multi-vector hybrid search, and scalar and JSON filtering with their own indexes. Version 3.0.0 was released on 29 July 2026.
It runs as the embedded Milvus Lite, a single Standalone server, or a Distributed cluster on Kubernetes. That last mode is where it fits teams leaving Pinecone because of scale, and also where the operational cost lies.
- Distributed mode depends on object storage, etcd and a WAL service (Kafka, Pulsar or Woodpecker), which needs operations skill.
- More index and tuning decisions than a managed service hides.
- The project has no managed service of its own; managed Milvus is Zilliz Cloud.
Zilliz Cloud
Zilliz Cloud is the commercial managed service built on Milvus. It offers Free, Serverless and Dedicated clusters on AWS, Google Cloud and Azure, plus Business Critical, On-demand Compute and BYOC deployments priced by sales. Serverless reads and writes are billed at $4 per million vCUs, and Dedicated clusters start from $126 a month on Standard and $197 a month on Enterprise.
For a team leaving Pinecone, the appeal is that the operational model stays similar while the engine underneath is Apache 2.0 Milvus, so a later move to self-hosting is possible without changing the data model (editorial).
- Index choice is reduced to AUTOINDEX, unlike self-hosted Milvus.
- Free clusters are limited to 5 GB and 5 collections and are suspended after 7 days of inactivity.
- Business Critical, On-demand Compute and BYOC are quote-only.
pgvector
pgvector is a PostgreSQL extension that adds vector types, distance operators and HNSW and IVFFlat indexes. Embeddings sit in ordinary tables, so you can join them to relational data, filter with SQL WHERE clauses and update them in the same transaction as the rows they describe. Version 0.8.7 was released on 1 October 2026.
This is the most common way out of a separate vector service: one database, one backup and one access model (editorial). Most managed PostgreSQL services offer it, though often a few releases behind (Amazon RDS listed 0.8.2 when 0.8.7 was current). See pgvector vs Pinecone.
- Indexes are limited to 2,000 dimensions for vector (4,000 for halfvec).
- Approximate index scans apply filters afterwards, so filtered queries can return fewer rows unless iterative scans are used.
- Hybrid search and embedding generation are not built in; you combine PostgreSQL full-text search and your own embedding code.
- Scale-out depends on PostgreSQL itself or extensions such as Citus.
MongoDB Atlas Vector Search
MongoDB Atlas includes Vector Search, which indexes vector fields in the same collections as the rest of your documents. Queries run as aggregation pipelines starting with a $vectorSearch stage; MongoDB documents HNSW for approximate search, exact search as an option, embeddings of up to 8,192 dimensions, pre-filtering on indexed fields, and hybrid queries that combine vector and full-text search. Automated Embedding can generate and keep embeddings in sync using a Voyage AI model.
MongoDB documents that Vector Search also works with self-managed or local Atlas deployments created with the Atlas CLI, and recommends separate search nodes for workload isolation in production.
- Vector Search is not supported on Atlas Infinite clusters during their public preview.
- Approximate search needs MongoDB 6.0.11, 7.0.2 or later on Atlas clusters.
- Dedicated Atlas costs grow with extra nodes, backups, data transfer and regions.
Elasticsearch
Elasticsearch stores dense vectors (up to 4,096 dimensions) and sparse vectors alongside text fields, with HNSW and optional int8, int4 or BBQ quantization. RRF and linear retrievers combine lexical and vector results in one request, and semantic_text fields can call the Elastic Inference Service, ELSER or third-party models, which covers much of what Pinecone's integrated inference does.
It is available self-managed, as Elastic Cloud Hosted (from 99 USD a month for a reference configuration) or as usage-based Elasticsearch Serverless. Free features are licensed under AGPLv3, SSPL or the Elastic License 2.0 at your option. See Elasticsearch vs OpenSearch for vector search.
- Some vector features, including the bbq_disk index type, need an Enterprise subscription.
- Self-managed clusters need shard, memory and node sizing.
- The licence history (Apache 2.0, then SSPL and ELv2, then AGPLv3 added) needs legal review if you redistribute it.
OpenSearch
OpenSearch is an Apache 2.0 search engine governed by the OpenSearch Software Foundation under the Linux Foundation since September 2024. Its k-NN vector field runs on the Faiss engine (the default, with HNSW and IVF) or Lucene, with JVector also available and NMSLIB deprecated. Hybrid queries with normalization or RRF have been available since version 2.11, and ingest pipelines can call registered or externally hosted models to create embeddings.
Amazon OpenSearch Service bills managed clusters by instance hour and storage with no minimum, or Serverless by OpenSearch Compute Unit hour; other providers such as Aiven also run it. It suits teams that want vector search under a permissive licence without adding a new kind of system.
- Engine options change over time (NMSLIB is deprecated), so older index mappings may need migration.
- Feature support differs by engine; check the engine matrix for the features you need.
- Managed AWS pricing has several dimensions (instances, storage tiers, OCUs) that make estimates harder.
Redis
Redis 8 includes the Redis Query Engine with FLAT, HNSW and SVS-VAMANA vector indexes, KNN and range queries with metadata filters, and FT.HYBRID for hybrid search since 8.4.0; vector sets are a separate data type added in 8.0.0. Because the data is held in memory, it suits caching-style workloads such as semantic caches and session-scoped retrieval (editorial).
Redis Open Source 8.0 and later is available under RSALv2, SSPLv1 or AGPLv3 at your option. Redis Cloud has a free 30 MB plan, Essentials from 0.007 USD per hour and Pro from 0.014 USD per hour with a 200 USD monthly minimum.
- Data set size is bounded by memory unless you use a tiering feature such as Redis Flex.
- Some SVS-VAMANA compression optimisations are Intel-proprietary and not in Redis Open Source.
- The licence changed twice since 2024; each option carries obligations to review.
When to stay with Pinecone
Stay with Pinecone if the main appeal was not running anything: none of the open source options removes operations work unless you buy their managed cloud, and the database-integrated options only help if you already run that database. Pinecone's free Starter plan and flat $20 Builder plan are cheap for prototypes, and Enterprise BYOC now answers some data residency questions without leaving the product.
Switching makes most sense when you need to self-host or avoid a closed source dependency (Qdrant, Weaviate, Milvus), want a managed service on an open engine (Zilliz Cloud), or would rather keep vectors next to the data they describe (pgvector, MongoDB Atlas, Redis, Elasticsearch, OpenSearch). Expect to re-embed or export vectors, map namespaces to the new tenancy model and rewrite query and filter code, and run both systems in parallel to compare result quality before cutting over.
Frequently asked questions
What is the best open source alternative to Pinecone?
There is no single answer. Qdrant (Apache 2.0) and Weaviate (BSD-3-Clause core) are standalone vector databases with managed clouds; Milvus (Apache 2.0) offers the most index types and a distributed mode. If your data is in PostgreSQL, pgvector avoids a second database.
Can I self-host Pinecone?
Not as a downloadable server. Pinecone is a proprietary managed service; Enterprise customers can run it in their own cloud account through BYOC, generally available since 25 August 2026.
Is pgvector enough instead of Pinecone?
For many applications whose data is already in PostgreSQL, yes, within its limits: HNSW indexes up to 2,000 dimensions for the vector type and scale-out through PostgreSQL itself. See pgvector vs Pinecone.
Which alternative has built-in embeddings like Pinecone?
Weaviate (model provider modules and Weaviate Embeddings), MongoDB Atlas (Automated Embedding with Voyage AI), Elasticsearch (semantic_text with inference endpoints) and Qdrant Cloud (Cloud Inference) can create embeddings for you. pgvector and Milvus self-hosted rely on your own code or, for Milvus, a Function module that calls external providers.
What about Vespa?
Vespa is an Apache 2.0 search, ranking and vector engine with a managed Vespa Cloud. It is powerful for combined text, vector and machine-learned ranking, but it has more concepts to learn than a vector API, so we have not listed it as a direct Pinecone replacement. See the Vespa profile.
Sources
- Pinecone pricing
- Pinecone 2026 release notes
- Pinecone: pod-based indexes (legacy)
- Qdrant pricing
- Qdrant hybrid queries
- Weaviate pricing
- Weaviate repository and licence
- Milvus index types
- Milvus release notes
- Zilliz Cloud pricing
- pgvector repository
- MongoDB Vector Search overview
- Elasticsearch kNN search
- Elastic licensing FAQ
- OpenSearch k-NN methods and engines
- Amazon OpenSearch Service pricing
- Redis vector search
- Redis licences
Checked November 2026.
How we research these guides: our editorial method.