Sberbank Releases Its First Open Neural Network With a Reasoning Mode
GigaChat 3.5 Reasoning delivers accuracy comparable to the world’s leading language models.
Neural networks are evolving at a rapid pace. They are no longer just electronic assistants that can quickly find information or solve a simple problem. Today, they are becoming tools that can analyze complex tasks and work through them step by step.
The Neural Network Fixes Its Own Mistakes
Sber has released a new version of its flagship neural network, called GigaChat 3.5 Reasoning. According to its developers, it is currently the only Russian open model that reasons through a problem before responding, works through chains of tasks and can correct its own mistakes.
When interacting with a user, GigaChat 3.5 Reasoning does not rush to answer difficult questions. Instead, the model breaks the task into steps and creates a plan for solving it. Only then does it turn to search and other tools. Before delivering the final answer, the system analyzes it again and corrects errors that may have occurred during processing.
The model is designed not only for speed but also to minimize the resources it uses. That gives it an advantage over open neural networks built around the traditional model format.
Built for Analysis
GigaChat 3.5 Reasoning is specifically tuned for analytical work. For example, the model can identify contradictions between clauses in a contract, perform financial calculations or troubleshoot a complex coding error.
Importantly, users still have a choice. They can turn the reasoning mode on or off with a separate button. In other words, if a user primarily needs quick answers, that option remains available.
With reasoning support, GigaChat’s performance has improved across almost all benchmarks. The most noticeable gains have come in agentic tasks, work with functions and tools, programming and problems that require complex logic. In key areas such as following complex instructions, planning and code generation, GigaChat 3.5 Reasoning has come close to the world’s leading models.
A reasoning mode could be useful across a wide range of fields, including law, finance, education, software development and many others. The new model is already available to all users. Developers can also use it free of charge under the MIT license to integrate it into their own products. An API version for businesses will become available soon.
Rapid Evolution in a Short Time
GigaChat has taken a major evolutionary leap in a relatively short period. Sber announced its launch in April 2023 and began actively integrating the neural network into its tools. By August of that year, it had been integrated into the Salyut voice assistant. Owners of SberBoom and SberBoom Mini smart speakers were the first to try out the service.
Over the next two years, neural networks gained traction in the corporate sector. At first, chatbots became their main area of application. But efforts to adapt AI for more complex business tasks continued. Within Sber’s ecosystem, solutions based on the GigaChat API were developed to help organizations create intelligent assistants.
In July this year, Sber introduced GigaChat 3.5 Ultra. According to its developers, the model generates text four times faster than the previous generation, requires fewer computing resources and is almost twice as compact. At the same time, the focus was on expanding capabilities that could be particularly useful in the corporate sector. GigaChat 3.5 Ultra, for example, gained improved capabilities for programming as well as mathematical and financial calculations. The model writes and checks code more accurately, handles numbers better, and produces answers that are more structured and easier to understand.
The release of GigaChat 3.5 Reasoning marks another step toward turning a neural network from a simple electronic assistant into a tool capable of handling genuinely complex tasks. In the coming years, competition in the AI market is likely to shift toward the quality of agentic systems: their ability to understand context, work with data specific to a user or company, keep information secure and execute complex workflows without constant human supervision.
