Milvus pricing

About

Milvus is the world’s most widely deployed open-source vector database, created by the engineers at Zilliz. The software is free under the Apache-2.0 license and is a graduated, top-level project of the Linux Foundation’s LF AI & Data Foundation — so the technology has no list price at all; self-hosters pay only for the compute, storage, and operational effort to run it.

Zilliz monetizes through Zilliz Cloud, the fully managed version of Milvus. This open-core split is the whole pricing story: the database is a commodity good given away to win developer mindshare, and the company captures revenue by removing the operational burden (cluster sizing, indexing, scaling, backups, compliance) of running vector search in production.

Zilliz raised a $60M Series B extension in August 2022 (on top of an earlier $43M), bringing total funding to roughly $113M, and relocated its headquarters to the San Francisco Bay Area to expand its engineering and go-to-market teams around the managed cloud.

As of July 2026 the managed offering is packaged as four project plans — Free, Standard, Enterprise and Business Critical — crossed with four deployment options: Serverless, Dedicated, the newly launched On-demand Compute (public preview, for lake-scale jobs on zero-copy external data), and BYOC.

For the most current information, visit Zilliz Cloud pricing and the Milvus project site.

Pricing summary : How Milvus’s pricing model works

There are two distinct cost worlds. Milvus OSS is free — your only costs are the cloud/server infrastructure and the engineering time to operate it. Zilliz Cloud (managed Milvus) is consumption-based and sold along two axes at once: a project plan (Free · Standard · Enterprise · Business Critical) that sets your support, SLA and security envelope, and a deployment option (Serverless · Dedicated · On-demand Compute · BYOC) that sets how compute is metered. Unit rates vary by plan, cloud region and cluster type.

• Free — $0: 5 GB storage, 2.5M vCUs per month included, up to 5 collections, on a shared Google Cloud environment.
• Standard — From $0/month (Serverless) and From $126/GB/month (Dedicated) as shown on the pricing page. Fully managed vector databases with core APIs, backup/restore and basic monitoring, encryption in transit and at rest; manual scaling to 32 CUs. 30-day free trial.
• Enterprise (marked Most Popular) — From $197/month (Dedicated): 99.95% uptime SLA, audit logs, SSO (SAML 2.0), granular RBAC, multi-replica and elastic scaling, private endpoint and VPC peering, 24/7/365 support. 30-day free trial.
• Business Critical — “Let’s Talk” (sales-quoted): global cluster with DR, 99.99% uptime SLA with multi-replica, CMEK and full-path in-transit encryption, HIPAA-eligible, 30-minute emergency response.
• BYOC (Bring Your Own Cloud) — deploy in your own cloud account with the same features as SaaS Dedicated clusters; Book a Demo (no public price).

Metering by deployment option:

• Serverless — pay-per-operation, billed in virtual compute units (vCUs) at $4 per million vCUs for reads and writes, plus storage. Inserts cost 0.25 vCU per KB, deletes 1 vCU per entity, and each read costs a minimum of 6 vCUs, scaling with data scanned and returned. Idle compute is free; storage still bills.
• Dedicated — Query CU Unit Price × Total Query CUs × Cluster Runtime, billed hourly. The docs’ worked example prices an Enterprise-plan, performance-optimized cluster in AWS us-east-1 at $0.248/hour per CU. Suspended clusters stop CU charges but keep billing storage. Headline cluster-type rates on the pricing page: performance-optimized from $63, capacity-optimized from $16, tiered-storage from $5 — all per million vectors / month.
• On-demand Compute (public preview, badged New!) — usage-based Query CU plus Indexing CU, billed per minute of active job runtime with no always-on compute; the docs example uses $0.41 / CU / hour for both in AWS us-west-2 on the Enterprise plan.
• Storage — Storage Unit Price × Data Size × Duration, billed hourly (backups daily), rate set by region and cluster type; docs examples use $0.025/GB per month for a performance-optimized dedicated cluster and $0.04/GB per month for managed volume storage.
• Add-on lines — storage requests, data transfer and audit logs are separately metered.

What makes this different: Milvus separates the metering for writes (data volume inserted/deleted) from reads (query work + data scanned/returned), lets Serverless drop idle compute to zero, and now adds a per-minute On-demand Compute mode for lake-scale jobs over zero-copy external data — three different meters for three different workload shapes under one plan ladder.

Pricing by product

Milvus (open source)

Zilliz Cloud (project plans)

Zilliz Cloud (metered usage)

Sales motions across products: PLG/self-serve for Milvus OSS, the Free plan, Serverless and Standard (30-day free trials, $200 free credits in the cost estimator, AWS/GCP/Azure Marketplace subscription); sales-led for Enterprise at scale, Business Critical and BYOC.

Hidden costs : What Milvus users actually pay

The real cost of Milvus depends almost entirely on whether you self-host or use the managed cloud:

• Self-hosted (OSS): the “free” software still costs you GPU/CPU instances, memory (HNSW indexes are memory-hungry), object storage, and — most underrated — engineering time. Community reports note teams running self-hosted Milvus for months before realizing ops cost more than a managed plan would have at their scale.
• Zilliz Cloud Serverless: the trap is read amplification — each read is a minimum of 6 vCUs and grows with data scanned and rows returned, so high query-per-second RAG workloads can run up vCU bills even with cheap storage.
• Storage and transfer: storage is now cheap — the docs’ worked examples price dedicated-cluster and backup storage at $0.025/GB per month (AWS us-east-1, performance-optimized) and managed volume storage at $0.04/GB per month — but data transfer and audit logs are separately billed line items, and Dedicated CU-hours accrue whether or not the cluster is queried.

Want to estimate your own Milvus / Zilliz Cloud bill? Use the Milvus pricing calculator to model costs from your read/write and storage patterns.

Pricing evolution : Milvus pricing history and changes

Cadence

Tracked range: 2023–present, reconstructed from Zilliz Cloud release notes, pricing-update posts and the live pricing page. Periods not listed above were verified stable (0 price changes, 0 SKU additions).

Notable changes

• 2023-06-11 — Zilliz Cloud reaches general availability with a tiered model (Free, Dedicated from ~$99/mo, Serverless) and consumption-based billing.
• 2024 — Serverless consolidates on a pay-per-operation model billed in vCUs ($4 per million), charging only for read/write operations plus storage.
• 2025-10 — Tiered storage rebuild cuts storage 87% ($0.30 → $0.04/GB/month), reduces compute pricing ~25%, and adds cross-region backups + a Business Critical plan for regulated industries.
• 2026-01-01 — Storage pricing standardized at $0.04/GB/month across AWS, Azure and Google Cloud.
• 2026-07-21 — The pricing page publishes entry prices for the paid plans for the first time in our tracking: Standard from $0/month (Serverless) or $126/GB/month (Dedicated), Enterprise from $197/month (Dedicated) with a Most Popular badge, and Business Critical as “Let’s Talk”. Dedicated cluster-type headline rates appear alongside them — performance-optimized from $63, capacity-optimized from $16, tiered-storage from $5, all per million 768-dim vectors per month. Nothing already published moved: Serverless stays at $4 per million vCUs (6-vCU read minimum), Dedicated at the docs’ $0.248/hour per CU, storage at $0.025–$0.04/GB/month, and the Free plan at 5 GB + 2.5M vCUs + 5 collections. This is packaging becoming legible, not a repricing.
• 2026-07-21 — On-demand Compute launches in public preview (badged New! on the pricing page): a third compute meter that bills Query CUs and Indexing CUs per minute of active job runtime for lake-scale search and system-managed index builds over zero-copy external data. Queue time and failed jobs are not billed, the minimum billing unit is one minute, and clusters auto-suspend on inactivity. The docs’ worked example prices both CU types at $0.41/CU/hour (AWS us-west-2, Enterprise plan) — roughly 65% above the $0.248/hour per-CU rate the Dedicated docs quote for an always-on cluster (different region, so not a like-for-like comparison, but the direction is clear). That gap is the price Zilliz puts on never provisioning capacity you don’t use.

What’s unique : Milvus’s distinctive pricing mechanics

1. Open-core, two-priced. The product literally has two prices: $0 for the OSS engine, and consumption-based for the managed cloud. The free software is the funnel; the managed cloud is the business.

2. Read/write split metering. Serverless meters writes by data volume (0.25 vCU/KB inserted, 1 vCU/delete) and reads by query work (minimum 6 vCUs, scaling with scan + rows returned) — a finer-grained value metric than the per-pod or per-vector models rivals use.

3. Zero-cost idle. Serverless charges nothing for idle compute — only storage — so a large but rarely-queried index is cheap to keep online. The October 2025 object-storage-backed tiered architecture (S3 with >90% local cache hit rates) is what makes the $0.04/GB storage line economically credible, and the July 2026 cluster-type rates put a number on the same trade — performance-optimized from $63, capacity-optimized from $16, tiered-storage from $5 per million vectors/month, a 12× spread bought entirely with latency. Compare how peers meter the same dimension on the vector storage pricing theme page.

4. Plan and deployment are priced on separate axes. Since July 2026 the page sells a support/SLA/security ladder (Free · Standard · Enterprise · Business Critical) and a metering mode (Serverless · Dedicated · On-demand Compute · BYOC), and the two multiply rather than nest. The same Dedicated cluster costs more on Enterprise than on Standard because the plan sets the CU unit price — so buyers are choosing a compliance envelope and a consumption curve in the same transaction, and an SSO or 99.95%-SLA requirement quietly reprices every CU-hour they will ever run.

5. A per-minute meter for job-shaped work. On-demand Compute (public preview) bills Query CUs and Indexing CUs by the minute of active runtime over zero-copy external data, with queue time and failed jobs excluded. It is a deliberate answer to the batch/analytical end of vector search — the workload that neither Serverless (per-operation, tuned for interactive reads) nor Dedicated (always-on, tuned for steady QPS) prices well — and it gives Zilliz three meters that map to three genuinely different workload shapes instead of one meter stretched across all of them.

Strengths & weaknesses

Billing UX : Milvus billing controls and transparency

• Cost estimator on the pricing page — an “Estimate Your Cost” calculator with Dedicated / Serverless tabs, cloud provider, region, pricing plan, cluster type, number of entities, vector dimensions, data type and an mmap toggle; it returns an estimated monthly total broken into Query CU and storage lines (the captured default returns $126.10 = $126.00 Query CU + $0.10 storage for 1 CU and 4.05 GB).
• $200 free credits — the estimator offers “Evaluate your cost with $200 free credits” for running a proof of concept before committing.
• 30-day free trials — Standard and Enterprise both start from a “Start 30-day Free Trial” button; the Free plan requires no credit card (“Start for Free”).
• Marketplace subscriptions — “Subscribe on Marketplace” toggles for AWS, GCP and Azure; per the billing docs, credits and Advance Pay can combine with either a credit card or a Marketplace subscription, but not both payment channels at once.
• Billing & usage console pages — documented Payment and Billing Overview, Set Up / Update Payment Method, Billing Profile, Understand Invoices, Manage Invoices, Separate Billing on Marketplace, Failed Payments and Organization Recovery, plus Monitor Billing Alerts.
• Cost analysis tooling — the docs expose Analyze Cost and Cost Optimization guides, with cost broken out by category (on-demand compute, dedicated cluster, serverless cluster, storage, storage requests, data transfer, audit logs); indexing CU costs roll up by database rather than by job.
• Suspend to stop compute — suspending a Dedicated cluster halts vector-database (CU) charges while storage keeps billing until you delete the cluster; on-demand clusters auto-suspend on inactivity.
• Annual commitment discount — per the pricing FAQ, “when you commit to an annual plan, you can get additional credits based on your usage.”

Strategic wins : Why Milvus’s pricing decisions worked

1. Give away the engine, sell the operations

By open-sourcing Milvus under Apache-2.0 and donating it to the LF AI & Data Foundation, Zilliz turned the database into a default choice for AI builders, then monetized the part developers actually dislike — running it. This is the classic open-core play applied to vector search. See usage-based pricing strategy for the underlying frameworks.

2. Lead the price-down race in vector storage

The October 2025 87% storage cut and ~25% compute reduction were a deliberate move to stay ahead as vector-DB pricing compresses. Re-architecting onto object storage with a hot cache let Zilliz cut the price while improving margins — a structural win, not just a discount. Related: how AI companies structure pricing and outcome-based pricing trends.

3. Pick value metrics that match the workload

Splitting metering into write-volume (vCU/KB) and read-work (vCU/query) ties the bill to the actual value-producing operations rather than to provisioned capacity. See choosing the right usage metric.

4. Put an entry price on every paid plan, then badge the one you want bought

The July 2026 page turned an opaque rate card into a shoppable ladder: Standard from $0/month, Enterprise from $197/month with a Most Popular badge, Business Critical as “Let’s Talk”. For a consumption product this is harder than it looks — there is no natural “price” when everything is metered — and Zilliz solved it by quoting the cheapest real configuration rather than hiding behind “contact us”. The badge does the rest of the work: it anchors the buyer on the SLA/SSO tier before they have modelled a single CU-hour, which is exactly where a pure-usage pricing vendor wants the conversation to start.

5. Add a meter when a workload doesn’t fit, instead of bending an existing one

Lake-scale index builds and batch search are bursty, long-running and indifferent to latency — priced on Serverless they would detonate the 6-vCU read minimum; priced on Dedicated they would pay for 24×7 capacity to run for 30 minutes. Rather than discount around the mismatch, Zilliz shipped On-demand Compute with its own per-minute Query/Indexing CU meter and excluded queue time and failed jobs from the bill. Charging a premium per CU-hour for that elasticity is the honest trade, and it keeps the existing meters clean. See choosing the right usage metric and why AI companies are moving off per-user licenses.

Areas to improve : Gaps in Milvus’s pricing approach

1. vCU forecasting is hard

The minimum-6-vCU-per-read rule plus scan/return scaling makes Serverless bills tough to predict for query-heavy RAG. A clearer per-query cost estimator up front would reduce bill-shock risk. See bill shock and cost unpredictability.

2. Too many metered dimensions — and July 2026 added a seventh

The docs’ own cost breakdown now runs to seven categories: on-demand compute, dedicated cluster, serverless cluster, storage, storage requests, data transfer and audit logs. Adding On-demand Compute solved a real workload mismatch, but it also means a single project can now accrue per-minute CUs, per-hour CUs, per-operation vCUs and per-GB storage simultaneously, and indexing CU costs roll up by database rather than by job — so you cannot attribute a spike to the job that caused it. A per-job indexing breakdown in the Usage view, plus a bundled “all-in” line for smaller customers, would cut the cognitive cost of adoption.

3. The “free” trap for self-hosters

Marketing leans on “free and open-source,” but the total cost of self-hosting (memory-hungry indexes + ops) catches teams off guard. More honest TCO guidance vs. the managed cloud would build trust.

4. The published entry prices aren’t comparable to each other

Publishing list prices was the right call, but the two Dedicated entry points are quoted in different units — Standard “from $126/GB/month”, Enterprise “from $197/month” — so a buyer cannot tell from the page whether moving up the ladder doubles their bill or barely moves it. The page’s own estimator returns $126.00 for a single Query CU on a Standard capacity-optimized cluster, which suggests the $126 figure tracks a compute unit rather than a gigabyte. Quoting both plans on the same denominator (per CU per month, at a stated cluster type and region) would remove the ambiguity, and publishing the On-demand Query/Indexing CU rates on the pricing page — rather than leaving $0.41/CU/hour to a docs worked example — would let buyers price the preview meter before adopting it.

Key takeaways

1. Milvus is free; Zilliz Cloud is the product. Never conflate the OSS engine’s $0 license with the cost of running vector search in production.
2. Usage-based done well rewards idle. Serverless charging nothing for idle compute is a genuine UBP feature, not a gimmick, and July 2026’s On-demand Compute extends the same promise to batch work by auto-suspending on inactivity and excluding queue time and failed jobs from the bill.
3. Storage is racing to zero. The 87% storage cut shows vector-DB economics shifting from RAM-resident indexes to object-storage-backed tiers.
4. One meter per workload shape beats one meter stretched thin. Zilliz now runs three compute meters — per-operation vCUs, per-hour CUs and per-minute on-demand CUs — because interactive, steady-state and batch vector workloads genuinely price differently; the cost is that a single project can hit all three at once, so invest in per-job attribution before you add the third meter, not after.
5. Open-core seeds the funnel. Free OSS + a generous managed free tier is how Zilliz competes with closed rivals like Pinecone on adoption.

UBP implications

1. Charge for value-producing operations, not provisioned capacity. Milvus’s vCU/CU-hour split is a clean example of metering reads and writes separately.
2. Make idle free to win spiky workloads — and charge a premium for the elasticity. Zero-cost idle Serverless compute is a strong UBP differentiator for unpredictable AI traffic, and On-demand Compute takes it further by billing per minute of active job runtime with queue time and failures excluded. The docs’ $0.41/CU/hour on-demand example against $0.248/hour for an always-on Dedicated CU shows the honest version of that trade: buyers pay a higher unit rate precisely because they buy far fewer units.
3. Re-architect to cut price and protect margin. The object-storage tiered redesign let Zilliz drop storage 87% without bleeding margin — pricing power follows architecture.

Sources

• Zilliz Cloud pricing (accessed 2026-07-21)
• Understand Cost — Zilliz Cloud docs (accessed 2026-07-21)
• Serverless Cluster Cost — Zilliz Cloud docs (accessed 2026-07-21)
• Dedicated Cluster Cost — Zilliz Cloud docs (accessed 2026-07-21)
• On-Demand Compute Cost — Zilliz Cloud docs (accessed 2026-07-21)
• Storage Cost — Zilliz Cloud docs (accessed 2026-07-21)
• Free Trials — Zilliz Cloud docs (accessed 2026-07-21)
• New in Zilliz Cloud: Tiered Storage, BC Plan, Pricing Changes (Oct 2025) (accessed 2026-06-09)
• Zilliz Cloud free tier (accessed 2026-06-09)

Bottom line

Milvus is the rare AI-infra entry whose core product is genuinely free: the open-source vector database carries no license cost, and Zilliz makes money by managing it. Zilliz Cloud’s pricing is real usage-based — vCUs for spiky Serverless workloads, CU-hours for steady Dedicated clusters, and since July 2026 per-minute Query and Indexing CUs for lake-scale batch jobs — with a generous free tier and an October 2025 re-architecture that cut storage 87% and compute ~25%. That same July 2026 update finally put entry prices on the ladder (Standard from $0/month Serverless or $126/GB/month Dedicated, Enterprise from $197/month, Business Critical quoted), which makes the offer shoppable without making it simple: seven metered cost categories, a per-read vCU minimum and two entry prices quoted in different units still make bills hard to forecast, and the “free” OSS path quietly bills you in engineering time. Browse the pricing blueprint for more fully-researched company profiles.

Want to compare Milvus against other vector databases like Pinecone, Weaviate, and Qdrant? Browse the pricing blueprint.

Pricing timeline : Major events on a vertical axis

Each milestone below corresponds to a public pricing change, product launch, or material adjustment. Major events use a filled marker; minor adjustments use a faded one.

Four-plan ladder published + On-demand Compute launches in public preview

Jul 2026

The pricing page now publishes a Free / Standard / Enterprise / Business Critical plan ladder with entry prices (Standard from $0/month Serverless or $126/GB/month Dedicated; Enterprise from $197/month Dedicated), and adds On-demand Compute — a per-minute Query CU + Indexing CU meter for lake-scale search over zero-copy external data.

Storage pricing standardized across clouds

Jan 2026

Storage standardized at $0.04/GB/month across AWS, Azure and Google Cloud.

Tiered storage, 87% storage price cut, Business Critical plan

Oct 2025

New object-storage-backed tiered storage drops storage from $0.30 to $0.04/GB/month, compute pricing cut ~25%, plus cross-region backups and a Business Critical plan for regulated industries.

Serverless pay-per-operation model matures

Jan 2024

Serverless clusters settle on a pay-per-operation model billed in vCUs ($4 per million vCUs), charging only for read/write operations plus storage; idle compute is free.

Zilliz Cloud GA + Serverless preview

Jun 2023

Managed Milvus reaches general availability with tiered plans (Free, Dedicated from ~$99/mo, Serverless) and a consumption-based model; storage priced around $0.30/GB/month.