Yandex Neuro Scale 2025 Product Digest

On September 24, we held our seventh annual conference — Yandex Neuro Scale 2025. This year, there were more than 50 presentations and seven thematic tracks: Infrastructure, DevOps, AI Studio, AI in action, Data, Security, and Cases. We’ll tell you about the main product announcements, technological updates, and other news.

Infrastructure and network

Introducing a new accessibility zone

We announced the launch of a new availability zone, which will go live in 2026. It is based on Yandex’s new data center with a capacity of more than 40 MW, located near the company’s existing data center in the Vladimir Region.

The proximity of the two zones makes it possible to ensure minimal data transfer delay (< 1 ms), with a total channel capacity between them of up to 25.6 TB/s. This is especially important for business-critical applications, in particular high-availability databases, that demand low network latency. The minimum delay in data transfer will be useful for banks, retailers, and other companies that value the speed and continuity of business processes: transactions, ticket bookings, and requests to internal databases. At the same time, the two zones are independent of each other technically and logically, ensuring a high level of fault-tolerance. The energy efficiency of the data center at 1.1 PUE is 27% lower than the global average.

On-premise Yandex Object Storage

Many companies, striving for flexibility and full control when building their infrastructure, face restrictions that prevent data from being stored in a public environment. These restrictions include strict legal and regulatory requirements, internal security policies, or specific requests for performance and network latency in an existing on-premise environment.

Yandex Object Storage can now be deployed directly on your local infrastructure. This solution allows you to meet all requirements while maintaining the key advantages of the cloud-based version: one of the best S3 API implementations on the Russian market, flexibility, geo-distribution of data, versioning, and scalability to exabytes.

Updating Yandex BareMetal

The service has new high-performance configurations, and the server leasing interface has been significantly simplified. We’ve also added new management and security features, network redundancy with MC‑LAG, and power redundancy.

One of the key changes in the service is the ability to lease ready-made or customized servers. At the same time, you can build the server you need not just based on existing components, but also request for special servers to be prepared for your specific needs.

New computing platforms

Client services require higher performance. We opened up access to new computing platforms that process data up to three times faster at a cost comparable to previous platforms. Up to 288 vCPUs and up to 1.7 TB of RAM are now available per virtual machine. The platforms are suitable for real-time applications — streaming services and payment systems — and to process large databases fast.

Artificial intelligence

Yandex AI Studio platform has been updated

We presented a new version of Yandex AI Studio, a platform for creating AI agents and applications. It includes Agent Atelier for assembling agents, Model Gallery with Yandex’s open-source and proprietary models, new APIs for dialogs and voice, and ready-made tools for integration with companies' external services and applications.

Agent Atelier and agent constructor

The main update is Agent Atelier, an “atelier” for digital workers where you can assemble text and voice agents for business tasks. Agents are configured not only via new APIs and SDKs, but also through the visual interface: you can select a model, connect tools (e.g., data search or CRM integration), and set rules of conduct and dialog logic. Workflows allow you to collect complex scenarios without code, and MCP Hub provides ready-made integrations with external systems and services in the Yandex ecosystem for business: YDB, Yandex Object Storage, Yandex Data Streams, Yandex Tracker, and others.

Model Gallery as a model showcase

The updated Yandex AI Studio offers a wide range of AI models, allowing you to choose the best option for any scenario, from RAG assistants to multimodal applications. The gallery includes basic models (YandexGPT 5.1 Pro, YandexGPT 5 Lite, YandexArt), specialized models (classifiers, embedders, rerankers), and open-source models (Qwen3 250B, Gemma 27B VL). Companies can also fine-tune models using their own data and use them in the Model Gallery. Almost any model can be deployed on dedicated resources with additional performance guarantees using the Dedicated Inference Endpoint tool.

AI Search for RAG scenarios

AI Search is a new smart search component that helps agents find relevant data and use it in responses. Three command modes are supported:

• vector search that reponds according to the “meaning”
• hybrid, combining vector and full-text methods
• web search for relevant information via the Yandex Search API.

With AI Search, agents can work with corporate documents, knowledge bases, and open sources to create informed and verifiable answers.

New APIs

Yandex AI Studio has new APIs for advanced development:

• Responses API — for building dialogs with memory,
• Realtime API — for voice agents without pauses,
• Vector Store API — for corporate search.

Production‑ready, out of the box

Yandex AI Studio is now a solution that combines the creation of agents, ready-made models, and services for production, all in one platform. Companies can now move from experimentation to implementation more quickly, and developers can create AI applications without unnecessary infrastructure complexity and the need to independently assemble pipelines from a large number of open-source components.

AI assistants in Yandex Cloud services

We launched AI assistants for working with our platform. They can help you set up cloud infrastructure, process alerts from information security systems, administer databases, analyze graphs, and perform other tasks. All you need to do is send a simply worded request to the assistant via the chat. AI assistants will remove up to 30% of routine tasks from clients and accelerate their completion:

• An AI assistant in the cloud console can help you determine what cloud resources and services are needed for a particular task. It can even understand common requests, like “help build the infrastructure for an online store,” select the right parameters and, if the user agrees with them, deploy and configure the necessary infrastructure itself or write detailed instructions.

• On the Yandex Security Deck platform, the AI assistant will answer questions about IP services, analyze security alerts, and give recommendations on how to optimize protection. In the event of an incident, the neural network will help you quickly understand how critical it is and what to do.

• In the Yandex SpeechSense service, the AI assistant can help analyze conversations in a stream and extract the necessary data from them: phone numbers, city, products that interested customers, and company reviews. You can upload a knowledge base in PDF format to your assistant, so it can analyze scripted voice chats and highlight agent faults or omissions.

• In Yandex WebSQL, for example, the AI assistant can automatically request information needed from Trino, ClickHouse^®, PostgreSQL, and other databases on request in natural language, e.g., “show the top 10 customers in terms of revenue in 2025.”

• In Yandex MetaData Hub, you can use Data Catalog to automatically mark up metadata to improve the speed and quality of database management.

In addition, the AI assistant in YDB can independently diagnose and optimize inefficient queries to improve data processing speed. This means that the services can be used even by data engineers without extensive experience in database administration. This is one of the world’s first database AI assistants.

Yandex SpeechKit has been updated

The speech recognition and synthesis service now offers the option of automatic logging and smart summarization. It converts audio recordings of meetings, calls, and lectures into structured minutes, and also highlights key points, tasks, and the people responsible, and divides them up by speakers.

This saves time on taking notes and ensures that all agreements are recorded. This functionality is available through the Yandex SpeechKit Playground API and interface, and also supports on-premises delivery for sensitive data.

Developer tools

The SourceCraft platform has been updated

We introduced a new version of the SourceCraft team development platform. The SourceCraft Code Assistant now has an AI agent mode. SourceCraft was the first solution in Russia with an AI agent integrated into the full software development cycle within one platform. The platform also includes new security tools, advanced code navigation, and integration with the Yandex Cloud platform for deploying applications:

• Developers can use SourceCraft Code Assistant for targeted assistance when writing code, or switch to agent mode to automate complex development scenarios. The AI agent speeds up the process from an idea to deploying software on Yandex Cloud. It can help you create a repository, write code, generate automated tests, check security, prepare a pull request, and launch deployment on the cloud platform with a single user request.

• The code assistant also works in basic mode with hints and code autocompletion. The tool can now track where code fragements are used repeatedly. When editing a function, the system identifies connections within the code, automatically finds all the places where it is used, and offers ready-made fixes. A developer can go from one place to another automatically without manual search.

• Cloud and CI/CD integration. SourceCraft fully integrates with our platform — when performing CI/CD tasks, you can access cloud services without manually configuring access keys. We’ve added support for GitHub Actions: when you transfer a project to SourceCraft, the platform automatically converts ready-made scripts into its own format.

• Code navigation and security. We have expanded support for languages for code navigation, including C# and Kotlin. The platform automatically finds confidential data in the code — passwords, API keys — at the writing stage, before changes are made to the production version.

SourceCraft is available for free, and new users can get grants for cloud services for testing purposes.

Update to Yandex Managed Service for Kubernetes^®

• Version 1.33 has been released. In it, sidecar containers are now publicly available (GA), and the beta version allows you to change container resources without reloading.

• For all supported versions (1.30—1.33), the node image was updated to Ubuntu^® 22.

• We added support for Workload Identity Federation.

YTsaurus 25.2 was released

Key updates include support for clusters with GPU hosts and a built-in dynamic table bundle management tool.

Monitoring

Update for Yandex Monitoring

The service has significantly expanded the ability to visualize charts. Users can now granularly customize the way lines are displayed, controlling their color, thickness, style, and name. The same settings can be applied both to individual lines and to entire groups filtered using regular expressions.

This opens up many use cases, e.g., you can color lines from different data centers in different colors or display thresholds in dotted lines for greater clarity:

We have added the ability to post links on dashboards to quickly access detailed information or documentation:

You can also now mark dashboards with labels, making it easier to search and navigate the listing. On the page with the list of dashboards, a new field allows you to search by name, description, labels, and other characteristics.

Yandex Observability Platform

We have begun collecting applications for access to closed testing of the platform, which allows you to quickly and easily get an unambiguous answer about the state of the operating systems at any time.

Security

New modules in Yandex Security Deck

The service for integrated security management has new modules:

• Kubernetes^® Security Posture Management (KSPM). It provides a single space for developers and IS teams to monitor security without slowing down development processes. The module detects and protects Kubernetes clusters, allows you to define security policies during deployment, and protect applications at runtime.

• CSPM (Cloud Security Posture Management). It allows you to automate the process of monitoring the security of the configurations of the client’s entire cloud infrastructure.

• Threat Detection. It automatically monitors threats and displays warnings in the console after Yandex Security Deck is connected.

The updates also affect the DSPM module for detecting and controlling confidential data. It is now integrated with Yandex 360, so users can scan data from Yandex Disk and find personal, payment, and other sensitive data. In the near future, a similar functionality will appear for Yandex Mail.

ML WAF at Yandex Smart Web Security

A new ML WAF module has been added to the service. This is an ML-based security technology that does not require configuration, works immediately when connected, and allows you to instantly block web attacks not covered by signature rules.

Recommendations for securely developing AI agents and multi-agent systems

We have compiled guidelines for reducing risks when developing and implementing LLM-based AI agents. Inside, there is an approach for developing a threat model when implementing agents and recommendations on how to reduce security risks when creating multi-agent systems.

Serverless computing

New Yandex Workflows features

We added integration with AI services and new automation features. New integration steps have appeared: uploading and vectorizing files in VectorStore (for RAG scenarios), as well as AI Studio Models and AI Studio Agent cubes for direct access to models and agents.

For writing and debugging templated parameters in jq, Playground has appeared, as well as an editor with syntax highlighting and the option to run expressions. In addition, we have improved the usability of Workflows. You can now run workflows on a scheduled basis (cron) and perform direct SQL queries against relational databases (YDB, MySQL^®, Postgres, ClickHouse^®) to add data to the context.

Updates to Yandex Cloud Postbox and Yandex Serverless Containers

Yandex Cloud Postbox has added graphical event statistics to track email delivery, openings, link clicks, and unsubscribes. Yandex Serverless Containers now supports asynchronous container calls and the option to run Docker^® within a container.

Data platform

Neuroanalyst in the Business service plan

Yandex DataLens 's AI agent is now available to all Business users with a 30-day trial period. Neuroanalyst is available in the service in chat format: it analyzes data on dashboards, finds insights, and helps form conclusions on queries in natural language.

The agent helps with creating reports: it writes calculated formulas and generates code for custom visualizations. This makes analytics accessible, allowing non-technical specialists to work with data, and speeds up the testing of business hypotheses by an average of 30%. Neuroanalyst is available both in the cloud and on premises.

YDB

Introducing Enterprise Manager

The management system for clusters and YDB databases aggregates databas statuses in one interface and simplifies administration. It has a built-in AI agent that automates key tasks: scheme optimization, monitoring, fixing problems, and improving query performance. Enterprise Manager is fully available for on-premise deployments.

Users of the Yandex Managed Service for YDB cloud service will have access to the Enterprise Manager version with the AI assistant adapted to work in a managed environment. It provides all the tools you need to manage databases. The control functions for clusters and hardware resources remain on Yandex Cloud’s side, as they are part of the managed service.

Analytical operation mode

Yandex Managed Service for YDB now provides access to the database’s analytical capabilities. YDB can now query large amounts of data, and the option to add and remove compute nodes allows you to flexibly manage the size of the database, adjusting it to current tasks. Built-in support for the Kafka^® API allows streaming data to be supplied to YDB using familiar tools, and the built-in data transfer mechanisms automatically store streaming data in tabular form for further processing.

The cost optimizer and the option to explicitly control the execution plan provide optimal query execution plans, while integration with Yandex DataLens and Apache Superset^™ makes it easy to build data marts for business analytics based on YDB.

Kafka protocol support within YDB Topics

YDB Topics, a message broker within YDB, now works with Kafka^® Connect^™, Apache Flink^®, Confluent Schema Registry, and other streaming frameworks. This was made possible by adding compacted topics, Kafka transactions, and Kafka-like reader balancing to Kafka API YDB Topics. This makes it easier to migrate: you only need to replace the Kafka connection endpoints with the YDB address in the application source code.

YDB MCP Server

YDB MCP Server allows you to connect an LLM model to YDB so that it can directly generate and execute database queries. The user can access YDB in natural language, and the model can generate SQL queries, process errors, and return results in an easily readable form. It’s enough to connect an MCP client and set up the connection according to the instructions in the documentation.

Marketplace

Two-phase purchase

Yandex Cloud Marketplace has a new option — two-phase purchase (“access on demand” mode). It allows publishers to set up a presale directly through the platform: customers submit a request and can clarify product details before launching, and publishers get a convenient tool for communication and providing access. All interaction takes place in the console, with status tracking and notifications for both parties.

Yandex Cloud Gwin

A new Kubernetes controller for integration with Yandex Application Load Balancer, Yandex Cloud Gwin, has been added to the Marketplace. This product allows you to manage cloud load balancers via standard Kubernetes resources (Ingress, Gateway API) and its own Custom Resource Definitions (CRD). Gwin automatically creates and sets up an Application Load Balancer on Yandex Cloud based on the Kubernetes configuration, providing traffic routing, SSL termination, backend health checks, logging, and autoscaling.