Open Source Milvus Vector Database v 2.6

1 | © Copyright 2025 Zilliz
1
Built for Scale, Designed to Reduce Costs | June 2025
Milvus 2.6 Overview

2
Milvus is an Open-Source Vector Database to
store, index, manage, and use massive number
of embedding vectors generated by deep
neural networks.
contributors
300
stars
35.4K
active pods
100M
forks
3.3K+
Milvus: The most widely-adopted vector database

3
3 | © Copyright 9/6/23 Zilliz
BUILT FOR AI OPEN SOURCE Performant at Scale
Why Milvus
Designed from the ground
up for vector search
Fully open source under
Apache 2.0, no vendor lock-in
Handles billions of vectors
with sub-10ms latency
Check Fully Managed Milvus at zilliz.com

4
Milvus 2.6 at a Glance
Lower Infra Costs
Boost Developer
Productivity
Uncompromised
Performance
Streamlined
Architecture

5
Smarter Infrastructure, Lower Bills
Tiered Storage (hot/cold
separation)
RabitQ 1-bit Quantization
Int8 Vector and HNSW Support
Milvus Storage V2

6
RaBitQ
RaBitQ is a binary quantization method based
on the geometric properties of
high-dimensional space.
Key advantages of RaBitQ
High search accuracy
Hardware-friendly (optimized with SIMD
Can be combined with Indexes like FastScaNN

7
Storage V2
Split Large and Small field
Vector Stored outside parquet
Use Page Stats to accelerate
Point Query
TODO
More Data Types: TEXT, BLOB
Golang and Java Reader

8
Built-In Tools Developers Love
✔ Data-In, Data-Out powered by cutting-edge embedding models.
✔ The Struct List Data Model
✔ Phrase Match
✔ Multi Language Tokenizer
✔ Add Field For Online Schema Evolution
✔ Query Sampling
✔ Time-Aware Decay Functions
✔ Refined TTL Strategy

9
Milvus Data Model - Struct List
Primary Key Partition Key Fixed Field Dynamic Field Vector Field ROW ID Timestamp
Reserved by the System
● Uniquely identifies
an entity
● Varchar or Int64
● User defined or
auto-generated
● Optional
● User defined or
auto-generated
● Optimizes search
by narrowing
queries to relevant
partitions
● Pre-defined
● Supported data
types: Numeric,
Varchar, JSON,
Array
● Roadmap support
for: Set,
Geolocation
● Supports optional
fields without
schema changes
● Stored as key-value
pairs
● Allows scalar
filtering on dynamic
fields
● Stored in a JSON
string
● Dense Float32,
Float16, BFloat16
● Sparse
● Binary
● 10 vector fields per
entity can be
defined
● Used as
Multi-Version
Concurrency
Control MVCC) and
concurrency
guarantee
The same data model for
for Milvus Lite, Standalone,
Distributed and Zilliz Cloud

10
Fast by Default, Scale on Demand
✔ JSON Shredding and JSON Index
✔ Ngram Index
✔ MinHash LSH Index for Faster Data Deduplication
✔ Async Pymilvus Client
✔ VDB Bench 1.0 released

11
Json Shredding
{"a":10,"b":"str1","d":"42", f:1}
{"a":20,"b":"str2","d":"43", f:2}
{"a":30,"b":"str3","e":"44"，f:3}
{"a":40,"b": 1,"d":"foo","e":"baz"}
{"a":50,"b": 2,"d":["23","24"]}
{"a":60,"b": 3,"d":{"e":"bar"},"e":"45"}
Performance Improvement 10310x
in our dynamic field test

12
Less Plumbing, More Building
✔ Streaming Node + WoodPecker = no Kafka or Pulsar
✔ Merge IndexNode and DataNode
✔ Merge All Coordinator into MixCoord
✔ CDC Support Bulk Insert and All DDLs
✔ APT/YUM install support

13
Stream Node - Stay Fresh Without Slowing Down

14
WoodPecker - Diskless WAL on S3
https://milvus.io/blog/we-replaced-kafka-pulsar-with-a-woodpecker-for-m
ilvus.md

15
Read.AI Scales Conversational Intelligence with Milvus
Read.AI uses Milvus as the backbone of its semantic search infrastructure to index and query narrative-rich
embeddings at enterprise scale, achieving sub-20ms retrieval latency for millions of users across diverse
communication channels.
We've got millions of monthly active users
and all of the underlying data when we're
trying to go find related content, find
updates to an action item, find
recommendations... All of that under the
covers is being powered by retrieving data
from Milvus
— Rob Williams, Co-Founder and CTO, Read.AI
Read.AI needed to organize and
search unstructured
communication data across
meetings, chats, emails, and
CRMs that lived in disconnected
silos. At enterprise scale, they
required support for billions of
records across millions of
tenants with sub-20ms latency.
Previous vector database
solutions failed due to poor
multi-tenancy support and
limited filtering capabilities.
CHALLENGES
FAISS lacked multi-tenancy
while Pinecone couldn't handle
their embeddings and filtering
requirements. Milvus stood out
for its ability to scale to millions
of users, deliver consistent
sub-20ms latency, support
hybrid search workflows, and
provide strong multenancy. The
responsive developer
community and support during
proof-of-concept sealed the
decision.
WHY ZILLIZ CLOUD
With Milvus, Read.AI achieved a
5× speedup in search across
multimodal data while
maintaining 20ms latency with
complex filtering. They
successfully migrated millions
of accounts into enterprise
namespaces and now power
unified search across all
channels. Milvus enables
proactive insight delivery before
users ask, driving retention and
supporting enterprise upsells.
RESULTS

16
Case Study | Autonomous Driving
BOSCH, a global leader in autonomous driving technologies, needed a scalable way to store and retrieve vast
amounts of rare, high-dimensional driving scenario data. They adopted Milvus to power similarity search
across billions of vectors.
“Milvus enabled us to search across
billions of driving situation data in
milliseconds, helping us scale AI
development while cutting costs. Itʼs an
essential part of our autonomy stack.ˮ
— Mr. Zhang, Principal Software Engineer, Bosch
Collecting and managing rare
“corner caseˮ scenarios was
slow, expensive, and difficult to
scale. Conventional databases
and manual labeling approaches
couldnʼt meet the performance
or efficiency demands of
BOSCHʼs AI development
workflows.
CHALLENGES
Milvus provided the flexibility to
index and retrieve
high-dimensional vectors with
sub-second latency at
billion-scale. Its support for
quantization, sharding, and
modular design allowed BOSCH
to efficiently scale their
infrastructure and reduce
operational complexity.
WHY Milvus
BOSCH reduced data collection
costs by 80% and storage costs
by nearly $1.4 million per year.
With Milvus, they achieved
millisecond-level search
performance and accelerated
the development of autonomous
driving systems.
RESULTS

17
Getting Started on Milvus 2.6
Milvus 2.6 Release Blog
https://milvus.io/blog/introduce-milvus-26-built-for-scale-designed-to-redu
ce-costs.md
Open Source Milvus: https://milvus.io/
Fully Managed Milvus: https://zilliz.com/
Milvus Discord: https://discord.com/invite/33mfvwep3J

Open Source Milvus Vector Database v 2.6

More Related Content

Similar to Open Source Milvus Vector Database v 2.6

More from Zilliz

Recently uploaded

Open Source Milvus Vector Database v 2.6