verifier.org

Milvus / Zilliz Cloud alternatives

19 tools we tested head to head against Milvus / Zilliz Cloud, ranked — and what each one actually does differently.

last reviewed 23 jul 2026 · from our best 20 vector databases ·list curated by Onur Ozcanxin

first — what you'd be leaving

Milvus / Zilliz Cloud ranks #3 of 20 in our vector databases testing. the credible answer at a billion vectors.

88/100

the widest index selection anywhere and genuine billion-scale — and version 2.6 removed the external message queue that made it painful to run.

why people look for an alternative
  • −heaviest self-host here despite 2.6's improvements
  • −over-engineered below roughly 10m vectors
  • −the vendor also maintains the main public benchmark

stay with Milvus / Zilliz Cloud if widest index selection: diskann, gpu cagra, rabitq, ivf family, sparse is the thing you care about most — nothing below beats it on that.

the short version
best alternativepgvectoralmost everyone — start here and migrate on a measured threshold, not a vendor's table92/100
advertisement
  1. 1

    pgvector

    #1 in vector databases · the correct default for most teams, and it quietly got good

    92/100

    verdictfree, permissive, runs on every managed postgres, and the only option here that can update an embedding and its parent row in one transaction.

    pgvector vs Milvus / Zilliz Cloud
     Milvus / Zilliz Cloudpgvector
    pricefree self-hosted; zilliz cloud billed per compute unitfree — PostgreSQL License
    free tieryesyes
    licenseApache-2.0PostgreSQL License — permissive
    deploymentself-host, lite, or zilliz cloudextension on any postgres
    index typeshnsw, ivf, diskann, gpu cagra, scann, rabitqhnsw, ivfflat, binary quantization
    hybrid searchfull, with sparse indexesvia postgres full-text; you write the fusion
    pricing modelcompute units; rates largely unpublishedfree — you pay for postgres

    switch foralmost everyone — start here and migrate on a measured threshold, not a vendor's table

    pros
    • +free and permissively licensed, on every managed postgres
    • +transactional consistency between embeddings and business rows
    • +0.8.0 iterative index scans fixed the filtered-recall flaw
    • +half-precision and binary quantization cut footprint dramatically
    cons
    • −index builds are slow and lock-heavy at large scale
    • −no native distributed sharding
    • −you own all the tuning yourself
    visit pgvector →full pgvector review →
  2. 2

    Qdrant

    #2 in vector databases · filtering built into the graph traversal, not bolted around it

    90/100

    verdictthe best dedicated option: apache-2.0 with no strings, the strongest filtered-search design in the category, and the easiest serious self-host.

    Qdrant vs Milvus / Zilliz Cloud
     Milvus / Zilliz CloudQdrant
    pricefree self-hosted; zilliz cloud billed per compute unitfree tier; cloud billed hourly on compute and storage
    free tieryesyes
    licenseApache-2.0Apache-2.0
    deploymentself-host, lite, or zilliz cloudself-host or managed cloud
    index typeshnsw, ivf, diskann, gpu cagra, scann, rabitqhnsw + scalar/binary/product/turboquant
    hybrid searchfull, with sparse indexessparse vectors, native
    pricing modelcompute units; rates largely unpublishedhourly compute + storage, rates unpublished

    switch forteams who genuinely need a dedicated vector database and want to self-host it

    pros
    • +filters applied inside hnsw traversal — best filtered search here
    • +apache-2.0 with no revenue caps or service restrictions
    • +four quantization modes including turboquant
    • +single rust binary — easiest serious self-host
    cons
    • −cloud per-unit rates are not published
    • −hybrid and private cloud are quote-only
    • −smaller enterprise support footprint than the incumbents
    visit Qdrant →full Qdrant review →
  3. 3

    Elasticsearch

    #4 in vector databases · the licence objection to using it for vectors is largely obsolete

    85/100

    verdictbest-in-class hybrid retrieval with transparent serverless pricing — and its dense and sparse vector code sits in the agpl-licensed tree, not the restricted one.

    Elasticsearch vs Milvus / Zilliz Cloud
     Milvus / Zilliz CloudElasticsearch
    pricefree self-hosted; zilliz cloud billed per compute unitserverless from $0.09 per search VCU-hour; storage $0.047/GB/month
    free tieryesyes
    licenseApache-2.0AGPLv3 / SSPLv1 / Elastic 2.0 triple
    deploymentself-host, lite, or zilliz cloudself-host, cloud, serverless
    index typeshnsw, ivf, diskann, gpu cagra, scann, rabitqlucene hnsw, int8/int4, bbq
    hybrid searchfull, with sparse indexesbest in class — bm25 + rrf
    pricing modelcompute units; rates largely unpublishedvcu-hours + storage per gb

    switch forteams already running elasticsearch, and anyone who needs real hybrid search

    pros
    • +vector search is available under agplv3, not licence-restricted
    • +best-in-class hybrid with real bm25 and rank fusion
    • +transparent serverless pricing with 50gb free on vector profiles
    • +better binary quantization and mature filtering
    cons
    • −jvm heap tuning and real operational weight
    • −licence structure is complex even where it's permissive
    • −the elser sparse model is separately restricted
    visit Elasticsearch →full Elasticsearch review →
    advertisement
  4. 4

    Weaviate

    #5 in vector databases · multi-tenancy and hybrid search, priced per million dimensions

    84/100

    verdictthe strongest combination of hybrid search and per-tenant isolation here — with a pricing model that punishes large embeddings in a way no competitor's does.

    Weaviate vs Milvus / Zilliz Cloud
     Milvus / Zilliz CloudWeaviate
    pricefree self-hosted; zilliz cloud billed per compute unit$45 / month minimum; vector dimensions from $0.00465 per million
    free tieryesyes
    licenseApache-2.0BSD-3-Clause
    deploymentself-host, lite, or zilliz cloudself-host, cloud, or byoc
    index typeshnsw, ivf, diskann, gpu cagra, scann, rabitqhnsw + product/binary/scalar quantization
    hybrid searchfull, with sparse indexesfirst-class bm25 + dense fusion
    pricing modelcompute units; rates largely unpublishedper million vector dimensions

    switch formulti-tenant saas rag where each customer needs isolation

    pros
    • +native multi-tenancy with isolated shards and cold offloading
    • +first-class hybrid search with configurable fusion
    • +bsd-3-clause — clean permissive licence
    • +excellent documentation
    cons
    • −dimension-based pricing doubles when you double embedding size
    • −$45/month minimum on the entry cloud tier
    • −historically memory-hungry without quantization
    visit Weaviate →full Weaviate review →
  5. 5

    Chroma

    #6 in vector databases · the best developer experience, and honest per-unit pricing

    82/100

    verdictpip install and go, with genuinely transparent pricing — the right call up to a few million vectors and the wrong one well before a hundred.

    Chroma vs Milvus / Zilliz Cloud
     Milvus / Zilliz CloudChroma
    pricefree self-hosted; zilliz cloud billed per compute unit$0 base with usage-based billing; $0.33 per GiB-month storage
    free tieryesyes
    licenseApache-2.0Apache-2.0
    deploymentself-host, lite, or zilliz cloudembedded, self-host, or cloud
    index typeshnsw, ivf, diskann, gpu cagra, scann, rabitqhnsw
    hybrid searchfull, with sparse indexesmetadata + full-text; weaker than rivals
    pricing modelcompute units; rates largely unpublishedper gib stored, written and queried

    switch forprototypes and small-to-mid production where developer speed matters most

    pros
    • +best developer experience in the category
    • +genuinely transparent per-unit pricing with no base fee
    • +apache-2.0
    • +usage-based billing starts at zero
    cons
    • −hnsw only — thinnest tuning surface here
    • −unproven above roughly 10m vectors
    • −local mode and cloud are different code paths
    visit Chroma →full Chroma review →
  6. 6

    AWS S3 Vectors

    #7 in vector databases · storage at six cents a gigabyte, and it reprices the floor of the market

    80/100

    verdictroughly five times cheaper per gigabyte than the premium managed options, and explicitly built for cost rather than latency — choose it deliberately, not by default.

    AWS S3 Vectors vs Milvus / Zilliz Cloud
     Milvus / Zilliz CloudAWS S3 Vectors
    pricefree self-hosted; zilliz cloud billed per compute unit$0.06 per GB per month storage
    free tieryesno
    licenseApache-2.0proprietary, managed only
    deploymentself-host, lite, or zilliz cloudaws managed
    index typeshnsw, ivf, diskann, gpu cagra, scann, rabitqnot documented
    hybrid searchfull, with sparse indexesnot a strength
    pricing modelcompute units; rates largely unpublishedper gb stored, per million queries

    switch forlarge, cold, infrequently queried corpora where sub-second is acceptable

    pros
    • +$0.06/GB/month — the cheapest credible storage here
    • +two billion vectors per index, 31 regions
    • +fully published, legible rate card
    • +first 512KB returned per query is free
    cons
    • −sub-second latency — wrong for interactive agents
    • −128KB minimum per PUT punishes unbatched writes ~21×
    • −index algorithm and internals undocumented
    visit AWS S3 Vectors →full AWS S3 Vectors review →
  7. 7

    turbopuffer

    #8 in vector databases · object-storage economics with the best namespace multi-tenancy

    78/100

    verdictthe architecture the rest of the category converged on, with genuinely good per-tenant isolation — behind unpublished unit rates and no self-host option.

    turbopuffer vs Milvus / Zilliz Cloud
     Milvus / Zilliz Cloudturbopuffer
    pricefree self-hosted; zilliz cloud billed per compute unit$16 / month minimum usage
    free tieryesno
    licenseApache-2.0proprietary, managed only
    deploymentself-host, lite, or zilliz cloudmanaged; byoc at enterprise
    index typeshnsw, ivf, diskann, gpu cagra, scann, rabitqnot fully documented
    hybrid searchfull, with sparse indexessupported
    pricing modelcompute units; rates largely unpublishedusage, above plan minimums

    switch formulti-tenant products with large, mostly-cold corpora

    pros
    • +excellent economics on large, mostly-cold corpora
    • +namespace-per-tenant isolation is genuinely well designed
    • +low $16/month entry minimum
    • +well-regarded engineering team and public writing
    cons
    • −per-unit rates are not published anywhere fetchable
    • −closed source with no self-host option
    • −no free tier, and higher tail latency by design
    visit turbopuffer →full turbopuffer review →
  8. 8

    OpenSearch

    #9 in vector databases · apache-2.0 with no licence anxiety, and serverless finally scales to zero

    76/100

    verdictthe safe institutional choice on aws — genuinely apache-2.0, with faiss and lucene engines, and the always-on cost floor finally removed.

    OpenSearch vs Milvus / Zilliz Cloud
     Milvus / Zilliz CloudOpenSearch
    pricefree self-hosted; zilliz cloud billed per compute unitfree self-hosted; serverless billed per OCU-hour
    free tieryesyes
    licenseApache-2.0Apache-2.0
    deploymentself-host, lite, or zilliz cloudself-host or aws managed/serverless
    index typeshnsw, ivf, diskann, gpu cagra, scann, rabitqhnsw, ivf, faiss + lucene, disk modes
    hybrid searchfull, with sparse indexessupported, mature
    pricing modelcompute units; rates largely unpublishedocu-hours; scales to zero on nextgen

    switch foraws-committed teams who want vector search without licence questions

    pros
    • +apache-2.0 throughout with no licence ambiguity
    • +faiss and lucene engines, plus disk-based and quantized modes
    • +serverless nextgen scales to zero after 10 minutes
    • +native on aws with full hybrid search
    cons
    • −vector collections can't share compute units with other workloads
    • −forked-project ergonomics lag elasticsearch
    • −compute-unit accounting is confusing
    visit OpenSearch →full OpenSearch review →
  9. 9

    Vespa

    #10 in vector databases · the most sophisticated ranking engine here, with the steepest learning curve

    75/100

    verdicta full search and ranking engine rather than a vector store — if ranking quality is your product, rank it second; if you just need rag, it's overkill.

    Vespa vs Milvus / Zilliz Cloud
     Milvus / Zilliz CloudVespa
    pricefree self-hosted; zilliz cloud billed per compute unit$0.05 per vCPU-hour on the startup plan
    free tieryesno
    licenseApache-2.0Apache-2.0
    deploymentself-host, lite, or zilliz cloudself-host or vespa cloud
    index typeshnsw, ivf, diskann, gpu cagra, scann, rabitqhnsw + tensor ranking framework
    hybrid searchfull, with sparse indexesbest in class, multi-phase ranking
    pricing modelcompute units; rates largely unpublishedper vcpu / gb / gpu hour

    switch forproducts where retrieval and ranking quality is the differentiator

    pros
    • +unmatched ranking and hybrid sophistication
    • +most transparent resource pricing in the category
    • +apache-2.0 and proven at very large scale
    • +unit prices decline as allocation grows
    cons
    • −steepest learning curve here by a distance
    • −$20,000/month minimum on the enterprise tier
    • −overkill for straightforward rag
    visit Vespa →full Vespa review →
  10. 10

    Pinecone

    #11 in vector databases · genuinely zero-ops, with the hardest pricing here to forecast

    73/100

    verdictthe zero-operations promise is real and worth money to some teams — but you cannot tune it, cannot self-host it, and cannot easily predict the bill.

    Pinecone vs Milvus / Zilliz Cloud
     Milvus / Zilliz CloudPinecone
    pricefree self-hosted; zilliz cloud billed per compute unitfree tier up to 2GB; standard from $50/month minimum
    free tieryesyes
    licenseApache-2.0proprietary, managed only
    deploymentself-host, lite, or zilliz cloudmanaged serverless only
    index typeshnsw, ivf, diskann, gpu cagra, scann, rabitqproprietary; parameters not exposed
    hybrid searchfull, with sparse indexesnative sparse-dense
    pricing modelcompute units; rates largely unpublishedread/write units + storage per gb

    switch forteams who will pay a premium to never think about infrastructure

    pros
    • +genuinely zero operational burden
    • +native sparse-dense hybrid and good namespace multi-tenancy
    • +2gb free tier
    • +strong enterprise compliance posture
    cons
    • −unit-based pricing is the hardest here to forecast
    • −hnsw parameters not exposed — no tuning, no recall/latency trade
    • −no self-host option at all
    visit Pinecone →full Pinecone review →
+ 9 more tested, not detailed here
we ranked 20 vector databases in total. the 9 that didn't make this page are written up in the full ranking →

how these were compared

every tool on this page went through the same test as Milvus / Zilliz Cloud — same tasks, same order, scored the same way. the comparison tables are the figures from that testing, not vendor spec sheets.

the vector databases test in full →
was this useful?