GigaCode

2025

Agent mode to automate user routine tasks

Agent mode has appeared in the GigaCode developer assistant. It is designed to automate the user's routine tasks and is available for free to everyone authorized on the GitVerse platform. Sberbank announced this on November 20, 2025. Unlike closed B2B solutions on the Russian market, agent mode integrates into popular IDEs, including the JetBrains family and GigaIDE, and allows you to test the tool without restrictions.

In agent mode, GigaCode turns from an assistant into an active digital developer. He gets a command execution environment, the ability to access the Internet and access tools to independently perform design and code tasks from start to finish. It is enough to formulate the task in text form, and the agent will find and open the necessary files, make changes to the code, execute console commands, run tests, assemble a project or issue Comitas in the Git repository. This approach reduces the time spent on day-to-day operations and allows developers to focus on more complex tasks.

Compared to Russian counterparts that use open source plugins and development environments like VS Code, the solution is implemented with deep integration into JetBrains IDE, a popular development environment for creating software in many programming languages. This opportunity is especially relevant for the Russian market, where most companies use this platform. Moreover, Sberbank became the first to introduce a SWE agent (developer agent) back in June 2025 in the GigaStudio web application creation tool. And now users also have access to support for subagents in the SWE agent system. The primary agent can call auxiliary experts with their own context to analyze, write, and validate code, providing more accurate and faster results.

{{quote 'author=said Andrey Belevtsev, Senior Vice President, Head of the Technological Development Unit of Sberbank. | Starting agent mode based on GigaCode is a step towards simplifying development processes and improving command efficiency. Previously, users had to manually attach files to the chat for analysis, now the AI agent independently finds, opens the necessary ones and, if necessary, creates files, fully understanding the project context. Thus, the tool provides complete automation of routine tasks, allowing developers to focus on more complex aspects of their work. Unlike most solutions on the market, which rely on the adaptation of open-source models and plugins and do not allow you to receive analytics and evaluation of effects out of the box, our agent mode is Sberbank's own development, created from scratch on our infrastructure. This approach ensures excellent scalability, deep integration with metric services and high reliability, which is especially important for large projects and teams,}}

In addition, GigaCode received a significant update in version 3.0. Now the AI assistant includes the latest version of the code completion model - GigaCode Inline 4.0 13B-A3B, which includes 13 billion parameters and demonstrates a significant increase in efficiency, according to internal tests. The technology is fully developed by Sberbank's team from scratch and is based on the MoE ("mixture of experts") architecture, which is special for code models, thanks to which the load is optimally distributed between specialized modules, which guarantees a high speed of issuing recommendations. The model instantly offers ready-made code fragments for Python, Java, JavaScript and other languages, allowing you to write programs faster and better. Unlike open source alternatives (for example, QwenCoder 2.5 and 3, Mellum), it is 2-3 times less likely to make syntax errors and significantly surpass them in key modes: short prompts and using context from other files.

Sberbank continues to develop an ecosystem of AI tools for developers. GigaCode's corporate customers will also get free updates and access to agent mode.

GigaCode 2.0

Sber introduced GigaCode 2.0, an updated AI assistant designed to speed up the software development process. This version offers improved algorithms, improved performance and an updated interface. On June 25, 2025, Andrey Belevtsev, Senior Vice President, Head of the Technological Development Unit of Sberbank, spoke about this.

GigaCode 2.0 is based on an ensemble of models, including GigaChat and specialized code models, which are further trained on updated datacets, and their power is significantly increased. Moreover, the system has improved performance by 4 times, which allows you to generate up to 40 prompts per second, speeding up the process of writing code.

Advanced models better take into account the context of the code and make clues more meaningful and applicable in real-world tasks, including working with large codebases. This version also significantly improves the quality of responses to the benchmark of the international standard LiveCodeBench, which evaluates the accuracy and applicability of the generated code, from 28 to 65%.

GigaCode 2.0 also includes RAG (Retrieval Augmented Generation) technology, which has been improved to take into account the project context. This provides more accurate and relevant responses and reduces errors such as outdated methods or syntactic flaws. The tool supports more than 35 programming languages, including Python, Java, JavaScript, Kotlin, and C++, making it versatile for various projects.

Шаблон:Quote 'author=said Andrey Belevtsev, Senior Vice President, Head of the Technological Development Unit of Sberbank.

The GigaCode plugin was installed more than 45 thousand times on the GitVerse code platform, and about 25 thousand active users use it monthly. The tool is used not only by developers, but also by specialists in other roles, including DevOps, system analysts, QA engineers and data engineers. For example, PlantUML support allows system analysts to generate charts, which extends the scope of GigaCode.

GigaCode 2.0 is available for free to all users on the GitVerse platform as SaaS, and companies have access to an on-premium version of the solution for deployment in the company's infrastructure.

Availability for GitVerse users

The GigaCode agent, an autonomous AI-based service that can become a virtual member of the development team and perform code review tasks, has become available to GitVerse users. This was announced on June 4, 2025 by the general director of SberTekha Maxim Tyatyushev. Read more here.

2024

Based on GigaView

Russian developers have become available GigaView - a tool based on the AI GigaCode assistant, which increases the efficiency of teamwork on software projects. This decision was announced on December 12, 2024 by the senior vice president, head of the Technological Development unit. Sberbank Andrey Belevtsev More. here

Availability for Business

AI assistant GigaCode is now available to businesses for implementation in their own IT infrastructure. Sberbank announced this on October 8, 2024.

By purchasing the local version of the AI developer assistant GigaCode, the business will be able to deploy it within its own IT infrastructure, providing an even higher level of security and reliability. GigaCode will help maintain autonomy and business continuity even in case of technical or operational emergencies.

The on-prem delivery format includes a complete set and enterprise level support. The package includes a flexible set of components that allows you to adapt the product to your individual business needs. Deployment is possible on a standard infrastructure using generally accepted technologies without the need for specialized skills from the support team.

GigaCode is also available as a cloud service (SaaS), providing instant access to the product without the need to deploy and maintain your own infrastructure.

The AI assistant generates the most likely code sequels and offers them to the developer right in the programming process. With the help of multifunctional chat, GigaCode helps the developer cope with typical tasks (documenting, creating tests, refactoring, etc.), and can also provide assistance on any issues related to software development, code explanation, and more. The assistant supports more than 35 popular programming languages ​ ​ and more than 10 software development environments. GigaCode develops in synergy with the GigaChat service in terms of the technologies, models, processes and approaches used. Under the GigaCode hood there are several neural network models trained on a large amount of source code and specializing in solving development problems.

GigaCode is part of the developer teamwork product ecosystem. The ecosystem includes the GigaIDE development environment, which contains more than 70 additional development, testing and administration tools. As well as GitVerse - an open platform for joint development, with which developers can publish open and closed source projects.

2023

Opening Early Access

Sberbank opened early access to a personal AI developer assistant - the GigaCode service. GigaCode will help you generate the most likely code continuation directly in the development environment in real time by code fragment. The bank announced this on December 27, 2023.

The December 2023 solution supports more than 15 programming languages ​ ​ (Java, Python, JavaScript, TypeScript, C/C + + and others), as well as compatibility with many popular editors and integrated development environments (IDEA, PyCharm, VSCode, Jupiter and others). Other features of the service include advanced codogeneration, semantic code search, search for some types of vulnerabilities, and much more.

GigaCode develops in synergy with the GigaChat service in terms of the technologies, models, processes and approaches used. Under the GigaCode hood are neural network models that are trained on large source code cases and specialize in certain development tasks.

{{quote 'author=said Andrey Belevtsev, Senior Vice President, Service Station, Head of the Technologies Unit of Sberbank. | As promised earlier, we have opened access to GigaCode to external developers. In Sberbank, several thousand specialists have been working with this service every day and our calculations show that it speeds up the code development process by 25%. I am sure that GigaCode will become a universal AI assistant for a huge number of IT specialists, who can save them from performing routine and similar operations, }}

Product Announcement

In early September 2023, it became known about the creation by Sberbank of an AI assistant for writing GigaCode^[1]. It is able to generate the most likely continuation of it by a fragment of code. The service supports Java, JavaScript, Python, TypeScript, C/C + + and others, as well as IDEA, PyCharm, VSCode and Jupiter development environments.

The name GigaCode refers to the GigaChat chatbot based on the NeONKA and Kandinsky neural network models. Sberbank says that both projects are developing in parallel. The credit institution clarified that by the beginning of September 2023, several thousand Sberbank specialists use the GigaChat service; in the coming months, GigaCode will be available for external developers.

Sber created an AI assistant for writing code

The head of the practice "Machine Learning and Artificial Intelligence" Axenix (ex-Accenture) Aleksei Sergeyev, in a conversation with Kommersant, noted that the speed at which AI can generate code is important, but now models can successfully and quickly create only a small part of it. According to him, the feature is suitable for writing individual sections of code, functions and methods.

Nikolai Sokornov, director of the software development department at Reksoft, believes that such services can speed up the development process, but then the threshold for entering the profession will increase.

Beginners are most often assigned routine tasks - for example, writing a piece of code on a model several times. Tools like Copilot eliminate the need for such actions, "he says.

It is possible that the AI model will repeat the existing code, but whether this will be considered a violation is a question, said Anastasia Skovpen, an intellectual property lawyer. It depends on the ability of the copyright holder, whether he will be able to prove that his work was in the training data and that it was illegally reproduced or reworked. So far, there have been no such cases.^[2]

Notes