Hopsworks 5.0
by Hopsworks
The AI Lakehouse for real-time ML, now with coding agents living inside the platform
Hopsworks 5.0 is a sovereign AI Lakehouse that folds a feature store, model registry, vector database and serving layer into one platform, and embeds Claude Code and Codex directly inside it. It is aimed at ML platform teams in regulated industries that need sub-millisecond feature serving without handing their data to a managed cloud vendor.
Hopsworks is the commercial platform from the Stockholm company of the same name, founded in 2017 by Jim Dowling and colleagues out of KTH Royal Institute of Technology, which shipped the first open-source feature store in December 2018 and now sells a full AI Lakehouse. Version 5.0 reached general availability on 30 June 2026 and its headline change is structural rather than incremental: a development container with a coding agent and terminal is embedded inside the platform itself, pre-loaded with Claude Code and Codex and able to run several agent sessions at once, so a developer describes the pipeline they want and the platform handles ingestion through to a deployed inference endpoint without switching tools. Underneath sits the part that made Hopsworks' reputation — a feature store whose online layer runs on RonDB, an in-memory fork of MySQL NDB Cluster, delivering sub-millisecond retrieval without a separately operated Redis tier, with Spark handling batch transformation and Flink handling streams. Release 5.0 added native SQL through a Trino query engine, dashboards via Apache Superset, native Apache Iceberg support alongside Delta and Hudi, ingestion from Databricks and Snowflake, column-level access control, Hugging Face model imports into the registry, automatic provenance linking training data to feature views and model versions, and a claimed 50% latency reduction on common platform tasks. Deployment covers SaaS, on-premises, hybrid and fully air-gapped. Named customers include Zalando, Ericsson, Saab, Paddy Power and America First Credit Union.
The ML platform team that needs online feature serving at real-time latency and wants more capability than open-source Feast without Tecton's compute-scaled bill — particularly in a regulated or air-gapped environment.
Sub-millisecond online feature retrieval from RonDB with batch, streaming, registry and serving in one platform, deployable inside the customer's own perimeter.
At a Glance
- Category
- Data & Analytics
- Pricing
- Freemium, Usage-based, Subscription, Contact for pricing
- Target Market
- Data Scientists, MLOps Engineers, Enterprise Developers, CTOs, Data Engineers
- Deployment
- Hybrid, Self-hosted, Open-source, Cloud-first, Edge-first
- Founded
- 2017
- Headquarters
- Stockholm, Sweden
Key Features
- ✓Embedded coding agents
A development container and terminal inside the platform pre-loaded with Claude Code and Codex, supporting concurrent agent sessions so pipeline work never leaves the environment.
- ✓RonDB-backed online feature store
The online layer runs on an in-memory MySQL NDB Cluster fork, giving sub-millisecond feature retrieval without operating a separate Redis cluster alongside it.
- ✓Batch and streaming as equals
Spark handles batch transformation and Flink handles streams, so real-time features are a first-class path rather than a bolt-on with a capability cliff.
- ✓Open-table-format lakehouse
Reads Apache Iceberg, Delta Lake and Hudi in place with no migration, and 5.0 added Trino for native SQL and Apache Superset for dashboards at scale.
- ✓Unified registry with automatic provenance
Model registry imports directly from Hugging Face and automatically tracks lineage linking training data, feature views and model versions for audit.
- ✓Sovereign deployment
Runs air-gapped, on-premises or hybrid as well as SaaS, with column-level access control for precise data sharing between teams inside one platform.
Capabilities
Use Cases
- •Real-time fraud and risk scoring
Serve fresh behavioural features to a model at request time with sub-millisecond lookup, which batch-only feature stores cannot support.
- •Retail and media recommendations
Customers such as Zalando compute features once and reuse them across many models, avoiding the training-serving skew that separate pipelines create.
- •Regulated on-premises ML
Defence and telecom customers including Saab and Ericsson run the full lifecycle inside their own perimeter where no managed feature store is permitted.
- •RAG and agentic inference on private data
Use the built-in vector database and model serving to ground LLM applications in enterprise data without exporting it to an external vector store.
- •Agent-assisted pipeline development
Describe a time-series forecasting pipeline and let the embedded Claude Code or Codex session scaffold ingestion, features and a deployed inference endpoint.
Ideal For
Best For
- ✓Real-time inference systems — fraud scoring, recommendations, dynamic pricing — where online feature lookup latency is a hard requirement
- ✓Regulated and sovereign environments in finance, telecom and defence that need on-premises or air-gapped deployment rather than a managed cloud feature store
- ✓Teams that want batch and streaming features treated as equal first-class citizens instead of hitting Feast's capability cliff on real-time use cases
- ✓Organisations consolidating feature store, model registry, experiment tracking, vector database and serving instead of stitching five tools together
- ✓Shops standardising on open table formats, since Iceberg, Delta and Hudi are read in place with no migration
Not Ideal For
- ✗Teams that need polished interfaces and comprehensive documentation — independent comparisons say complex cases such as multi-entity features or time-windowed aggregations over high-cardinality keys send you to forums and Slack
- ✗Organisations without Kubernetes operational capacity for the self-hosted path; the Helm charts work but running them is entirely the customer's problem
- ✗Small teams that want a minimal feature store footprint, where Feast's simplicity is a better match than a full lakehouse
- ✗Anyone counting on a large community for answers — the open-source project has roughly 1,300 GitHub stars against Feast's much larger following
Integrations
Deployment
Market Analysis
Pros
- ✓Genuinely strong real-time serving: RonDB removes the separate online-store cluster most feature stores require
- ✓Streaming and batch are both first-class, avoiding what independent comparisons call the sharp capability cliff Feast has for real-time use cases
- ✓One platform covers feature store, model registry, experiment tracking, vector database and serving, with automatic feature-to-model lineage
- ✓Sovereign deployment including fully air-gapped, with named production customers in defence, telecom and financial services
- ✓Priced substantially below Tecton for comparable capability, with a real free tier that needs no credit card
Cons
- ✗Independent comparisons single out documentation gaps: multi-entity features and time-windowed aggregations over high-cardinality keys push you into forums and Slack
- ✗Self-hosting demands real Kubernetes expertise — the Helm charts work but operating them is entirely the customer's responsibility
- ✗A much smaller community than Feast, so there are fewer answers available when undocumented behaviour bites
- ✗The main open-source repository carries roughly 1,300 stars and was last pushed in February 2025, so the public repo is not a reliable signal of current development pace
- ✗No per-unit SaaS rate is published and on-premises deployment is Enterprise-only, so cost cannot be modelled before contacting sales
- ✗No retrievable G2, Capterra, TrustRadius or PeerSpot rating, and no meaningful Hacker News or Reddit thread, so there is no independent satisfaction score to cite
Pricing
Free
$0
- ✓1 project included
- ✓Feature Store and Model Registry
- ✓Community support
- ✓No credit card required
SaaS (Pay-as-you-go)
Usage-based
- ✓Unlimited projects
- ✓Feature Store, Model Registry and Model Serving
- ✓Platform SLA
- ✓Community support
- ✓Pay only for what you use
Enterprise
Contact for pricing
- ✓On-premises and air-gapped deployment
- ✓Dedicated support team
- ✓Custom integrations
- ✓Guaranteed SLA
- ✓GPU and distributed training
Hopsworks publishes tier structure but not rates: a genuinely free tier gives one project with the feature store and model registry and needs no credit card, the SaaS tier is pay-as-you-go on consumption with unlimited projects and a platform SLA, and Enterprise — the only route to on-premises or air-gapped deployment, dedicated support and a guaranteed SLA — requires a demo and a quote. No per-unit price is stated for the SaaS tier, so real spend cannot be modelled from the site. The commercially relevant point from independent comparisons is relative rather than absolute: Hopsworks sits substantially below Tecton's compute-scaled pricing while offering materially more than open-source Feast, which is precisely the middle ground it is sold into. Self-hosting the AGPL-3.0 open-source distribution carries no licence cost but the full Kubernetes operational burden.
Security & Compliance
Connect
Sources
This page was written from 6 sources, 3 on domains other than hopsworks.ai.
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
Hex
Agentic analytics platform: AI notebooks for data teams and conversational Threads for everyone, grounded in governed semantic context
Unstructured
ETL for unstructured data: turn PDFs, tables, emails and scans into AI-ready JSON for RAG and agents
Tacnode Context Lake
PostgreSQL-compatible context lake giving AI agents shared, live, transactional state at decision time
Reducto
Agentic document parsing API that turns complex PDFs, scans and spreadsheets into citation-grounded data for AI, now with the 1¢-per-page r-1 model