Sber launches GigaChat 3.5 Ultra as open source AI

Sber has launched its flagship artificial intelligence model, GigaChat 3.5 Ultra. It is available free through the company's GigaChat assistant and as open source for developers.

The release is Sber's latest effort to expand access to its in-house AI systems for consumer and commercial use. It is positioning the model as a tool for coding, mathematics, long-document analysis and agent-based tasks that carry out multi-step assignments with limited human input.

According to Sber, the new model writes and verifies code more reliably than its previous flagship system and produces long-form text up to four times faster. It also uses fewer computing resources and is nearly half the size of the earlier GigaChat Ultra model.

That smaller footprint lowers the hardware requirements for running the model independently. This should make the system more accessible to companies and developers that want to integrate it into their own software or use it to build AI agents.

Model design

The model is based on a proprietary architecture developed by Sber's team and uses linear attention technology. According to the company, this approach allows the system to retain the core context of long texts without repeatedly processing the entire preceding sequence whenever a new word is added.

In practical terms, the design improves the handling of contracts, technical regulations, reports and other lengthy documents. Sber said the model can analyse such material without losing context or accuracy, and at materially higher speed than its predecessor.

GigaChat 3.5 Ultra also uses a Mixture of Experts, or MoE, architecture. That structure is widely used in large AI models to assign different parts of the network to different tasks, improving efficiency compared with relying on a single dense model for every request.

Performance claims

Sber said internal tests showed the model outperformed its previous flagship in programming tasks, mathematical problems, multi-step assignments and Russian-language dialogue. It also said the model came close on several measures to results from open models such as DeepSeek 3.2, while being significantly more compact.

Sber did not provide a full benchmark table in the announcement, but said the latest version was designed for practical work rather than conversational use alone. That includes tasks in which the model is expected to search for information, write and execute code, call external services and return a completed result.

Those functions place GigaChat 3.5 Ultra in the growing category of so-called agentic AI systems. Businesses are increasingly testing such models for routine monitoring, data handling and scheduled reporting, although wider use still depends on reliability, cost and oversight.

Open access

Sber said the model is available to ordinary users through its GigaChat AI assistant for personal and workplace tasks. It has also released the system as open source, making it available to developers globally for use in their own products and research.

Open source is becoming a more contested part of the AI market as companies seek influence beyond their domestic user bases. By releasing a large model with its own architecture, Sber is attempting to build a broader developer ecosystem around technology developed internally rather than relying on Western platforms.

Sber said GigaChat 3.5 Ultra is one of the largest open-source models built with linear attention. It added that the model was trained on an expanded dataset focused on natural, human-generated texts that went through multiple stages of classification and filtering.

Anton Frolov set out Sber's view of the release in a statement accompanying the launch.

"We are living in a time when the gap between human capabilities and AI potential is shrinking rapidly. GigaChat 3.5 Ultra is our step toward what an AI tool for real-world tasks should be: a full-fledged partner capable of thinking within the logic of a specific process, not just answering questions. To develop such a model, you need to constantly experiment and try things no one has done before - the number of our experiments has more than doubled, reaching 1,500. We have proven that it is possible to build a strong model using a proprietary architecture and with fundamentally fewer resources. We want our solutions to become the foundation for new products and research that go far beyond Sber," said Anton Frolov, Senior Vice President, Head of GenAI Development, Sber.

Sber remains Russia's largest bank, and the launch shows how major financial groups continue to invest in proprietary AI models alongside their core banking operations. According to the company, GigaChat 3.5 Ultra was built entirely by its own team using a domestic architecture based on linear attention technology.