YandexGPT

The main articles are:

• Chatbots (Chat-bot) Virtual interlocutors
• Speech Recognition (Technology, Market)
• Speech technology: On the path from recognition to understanding
• Neural networks (neural networks)

2025

Possibility of further training by LoRA

Yandex B2B Tech has provided customers with the opportunity to further study neural networks using the LoRA (Low-Rank Adaptation) method. Now companies will be able to quickly and independently adapt neural networks for specialized tasks: for example, train the model to write corporate-style letters or extract important data from documents. The use of a trained model will be no more expensive than the basic versions of neural networks. This method is already available to all customers on the Yandex Cloud AI Studio AI development platform. The company announced this on February 27, 2025.

For further training by the LoRA method, the YandexGPT neural network and the LLaMA open source model are already available. This will soon be possible for models that can be launched on demand at dedicated capacities - for example, Qwen and DeepSeek. Thus, the business will be able to use several advanced neural networks in its solutions, easily switch between them and combine them for different scenarios.

Further training of the model allows you to improve the quality of solving specific problems: after customization, neural networks better summarize and more accurately classify texts, they more accurately follow the given response format and adhere to a certain role in the dialogue. The quality of answers to these problems, even on lightweight models, may be higher than in powerful versions of neural networks. Additional training is especially useful when the requirements for an answer are difficult to formulate in a short and understandable promta.

LoRA works quickly due to the fact that it changes a small number of parameters during additional training of the model. The algorithm focuses on the reference queries and responses pre-defined by the user and brings the results of the model closer to the samples from the datacet. When processing small data sets, the entire further training process using LoRA can take only 10 minutes.

The LoRA algorithm has already been recognized as one of the most effective methods of further training, and it is widely used to customize models among ML developers, including Yandex. The introduction of such a tool is another step towards the diversity and adaptability of technologies that simplify the solution of client problems, "said Grigory Atrepyev, CPO Yandex Cloud.

Additional training of models is especially relevant for medium and large businesses to optimize the most frequent business problems that are solved using neural networks. As part of technical testing, LoRA has already tried more than 100 customers: they used it to develop Telegram bots for, retail classify support requests, generate advertizing content and other tasks. In particular, Presentsimple.ai uses the YandexGPT 4 Lite model, which was further trained using LoRA on the results of the Pro version, to automatically generate presentations at the request of the client. According to the company, in this way the cost of generating presentations decreased six times, and thanks to the high speed of the model, they were able to create a series of volumetric slides for customers in 30-40 seconds. The completed model better adhered to the set technical task in terms of the number of characters, sentences and style.

Open access YandexGPT 5 Lite Pretrain

Yandex has published in the public domain YandexGPT 5 Lite Pretrain - a large language model (LLM) with 8 billion parameters with context support of 32 thousand tokens. This model surpasses the quality of responses of foreign analogues in most benchmarks. It will be useful for developers and business representatives who can further train it for their tasks and not waste time creating their own base model. The company announced this on February 26, 2025.

The pretrain version of the model already has extensive knowledge of the world, understands the context and patterns of the language. She was trained using Yandex technologies in two stages. At the first stage, the model was trained on Russian-language and English-language texts, code, calculations and other data with a total volume of 15 trillion tokens - this is like 2.5 billion web pages. At the second stage, high-quality data with a volume of 320 billion tokens was used, including texts of educational topics. To enrich the knowledge of the model, Yandex has improved the method of selecting data in order to use more useful text materials from the Internet, without losing their quality.

As of February 2025, more than 120 open Yandex technologies in this area are available to developers. The Basic version of the model can be adapted to suit any needs. It has the widest possible application and can be used in development, business and scientific research. In addition, thanks to its compactness, YandexGPT 5 Lite Pretrain does not require large computing power and will be useful for those who need a small Russian-language model.

Compared to the previous generation model, YandexGPT 4 Lite Pretrain, this model shows a significant increase in quality in solving mathematical problems and writing code. And compared to similar foreign models such as LLaMa3.1-8B and Qwen-2.5-7B-base, YandexGPT 5 Lite Pretrain leads in almost all types of tasks.

The last time Yandex posted a large language model in open access in 2022 was YaLM-100B. At the same time, Yandex remains among the leaders in the number of open source solutions in the field of machine learning (ML) and working with data according to ITMO research. In particular, the company continued to share tools for working with LLM - such as the YaFSDP library, which accelerates the training of large language models by up to 25%.

YandexGPT 5 Lite Pretrain is one of the new generation of Yandex neural networks - YandexGPT 5 - and is available for free on HuggingFace. The second model of the family, YandexGPT 5 Pro, is available to business representatives on the Yandex Cloud cloud platform, and to users in a chat with Alice.

For YandexGPT 5 Pro training, Yandex added more complex and diverse examples to the dataset, complicated tasks, added methods of training with DPO (Direct Preference Optimization) and PPO (Proximal Policy Optimization) reinforcement, as well as its modification - LogDPO, which increases the stability of model training. The company optimized the training process using the YaFSDP library, which reduced the need for computing resources by 25%.

YandexGPT 5 Pro Launch

The new language model YandexGPT 5 Pro is ahead of international counterparts in a number of parameters, including GPT-4o from. OpenAI This was Yandex reported in "" at the end of February 2025.

By this date, the neural network is already available for integration through the Yandex Cloud platform and is focused on a wide range of business tasks, including client service automation, text generation and data processing.

The model does not yet support multimodality (external communication through audio, photo or video), does not support file processing until only text.

YandexGPT 5, as stated, is a competitive model, especially for tasks in Russian, with strong integration into popular services and access for business.

According to the results of an internal blind pair comparison, YandexGPT 5 Pro in 67% of cases exceeds the previous version of YandexGPT 4 Pro, in 54% of cases - GPT-4o and in 64% of cases - the Chinese model Qwen2.5 32b when solving standard problems.

The Russian development demonstrates a particularly noticeable advantage in the tasks of classifying information (70% of victories against GPT-4o), extracting information from the text (71% versus 48%) and summarization (74% versus 56%). In the field of text reformulation, YandexGPT 5 Pro is inferior to Qwen2.5 (58% versus 66%), but still surpasses GPT-4o (51%).

It is often important for business customers that the model responds based on specific sources - for example, the company's knowledge base. This technology is called RAG (Retrieval Augmented Generation). In working with such scenarios, YandexGPT Pro 5 wins over both the previous generation model and the GPT-4o.

It turned out to increase the reliability of answers based on the correct sources by 71% and reduce the number of "hallucinations" - cases when a neural network invents facts.

YandexGPT 5 Pro also supports structured output mode, which allows you to receive responses in a strictly defined format according to the specified JSON scheme. This greatly simplifies the automatic processing of information, for example, when filling out product cards on marketplaces or creating AI agents.

In addition, the model received improved support for functions calling - special commands for integration with external tools and APIs. Developers can specify a list of available functions and parameters, and the neural network will suggest which call needs to be made to solve a specific problem.^[1]

YandexGPT 5 Pro is available in:

• Chat with Alice (in the web version, mobile application and Yandex Browser).
• Smart devices with Alice.
• Yandex Cloud AI Studio and SpeechSense corporate services.

The free limit is 20 messages per day, paid access is only 100 rubles per month.

In general, in the tests presented by Yandex, YandexGPT 5 Pro plus or minus is approximately comparable to GPT-4o and Qwen 2.5 32b, but it copes better with information in Russian, because it studied at Russian-language sources and is optimized for Russia - this is the main advantage.

2024

Implementing Mango Office in Speech Analytics

MANGO OFFICE and the Yandex Cloud cloud platform have expanded their technology partnership. Now MANGO OFFICE products use the YandexGPT generative language model. Read more here

Implementation at Universe DG

The company Universe Date"" has built neuronet YandexGPT into the software product for organizing processes. data management Universe DG Artificial intelligence allows users to generate descriptions for business terms and automate the creation of - SQL quests. This was DIS Group announced on September 30, 2024. More. here

Availability on the ValueAI platform

ValueAI Support feature 3 is available on the platform YandexGPT. Now companies can fully use the potential of the generative text model Yandex to solve complex business and work tasks with. This was data WaveAccess (WaveAxess) announced on September 18, 2024. More. here

The first collection of stories written by a neural network in Russia has been published. Used YandexGPT platform

In September 2024, the first collection of short stories written by the YandexGPT neural network in collaboration with modern Russian writers was published in Russia. The collection, entitled "Mechanical Intervention," includes 15 stories, each created with the participation of artificial intelligence. The Alpina Publisher publishing house presented the book at the Yandex Museum, where an official presentation was organized. Read more here.

Integration with Stream Evaluation 360

in YandexGPT integrated Stream Score 360"" and accelerated the development of employees four times. The company "Stream" announced this on July 24, 2024. More. here

Availability in Flow Engagement

On July 1, 2024, the HRTech company TalentTech announced the launch of the YandexGPT generative network in the Stream Engagement product. Read more here.

Availability in Yandex.Translator

On June 7, 2024, Yandex presented an updated version of machine translation, trained using YandexGPT. The large language model YandexGPT has prepared reference examples of texts on which the neural network in Yandex Translator was trained. This allowed the service to more accurately understand the context, recognize phraseologisms and navigate professional vocabulary. Read more here.

Availability in Yandex.Directive

Entrepreneurs who do not have a site will be able to advertise their goods and services in Yandex.Directive using neural networks. Yandex announced this on April 24, 2024. YandexGPT and other Yandex neural networks will create texts for the advertising campaign. Read more here.

Introduction into the "Plus7 MayK" system

The company Innovation date"" has built neuronet YandexGPT into the system for centralized knowledge management. Plus7 MayK Thanks to this, users directly in the system interface can make requests and process text information using generative. artificial intelligence This "" was Innovation date reported on April 22, 2024. More. here

Integration into SkillStaff platform

AWG (ArtWeb Group) has added a revision to the SkillStaff platform based on the YandexGPT API. The generative neural network creates a generalized description of the resume of IT specialists on the marketplace and helps customers save time when selecting specialists. AWG announced this on April 9, 2024. Read more here.

2023

YandexGPT application in Yandex Real Estate

Yandex.Real Estate has simplified the search for an apartment using the YandexGPT generative neural network. Yandex announced this on December 19, 2023. Read more here.

The ability to identify errors in the code and correct them

The YandexGPT neural network now detects errors in the code and helps fix them. The function is available for Yandex Practice users and the USE in Informatics with Yandex Textbook platform. The update will make learning more comfortable for users of both services and give them the opportunity to master new skills faster. It will also remove some of the routine tasks from faculty, help motivate students and students to continue learning and free up time and resources to immerse themselves in the profession and communicate with mentors. Yandex announced this on December 7, 2023.

Thanks to the update, service users will be able to understand why the code does not run and how to fix it. The neural network indicates typos, references to non-existent elements, errors in syntax, control and other components of the code. At the same time, it does not give ready-made solutions, but helps the user find the error and fix it.

According to internal testing, more than 70% of the students of the Workshop, having made a mistake, turn to the neural network for a hint. Helping to find and explain errors is especially important for students who have just started learning programming and have difficulty debugging code. This is what often stops them from continuing their studies. One of the requirements for the model is that it should not give the student a ready answer. Instead, she offers several hypotheses for the occurrence of an error, "said David Roganov, technical director of Yandex Workshop.

In Yandex Practice, a smart code editor is available in test mode in online course simulators in the following areas: front-end development, product design, full-stack development, Go-development from scratch, automation of testing in Python and Java, etc. The neural network produces three hypotheses, and if they do not fit, you can ask it to explain the error differently.

For Yandex Textbook users, the function is available within the framework of the USE in Informatics with Yandex Textbook project. The "Help fix the error" button will appear automatically in the AI assistant mode if the code is written incorrectly.

Training a neural network to detect errors in the code is a rather difficult task, especially when it comes to code written by children. When configuring a neural network, it is important to predict which incorrect decisions she has to correct, and schoolchildren write code with unpredictable errors. Our neural network is trained on many children's solutions - both correct and wrong. To collect this data, Yandex Textbook analyzed more than 50 thousand solutions to USE tasks from children from all over the country, collected as part of the Code Error project, - said Roman Levin, senior methodologist of Yandex Textbook.

Workshop and Textbook plan to continue to introduce artificial intelligence into the educational process. The workshop develops a chat that will help students with homework, errors in the finished code and ideas for the project. Yandex Textbook is working on the creation of the first educational neural network in Russia for the study of computer science.

Access to YandexGPT API for all users

On December 6, 2023, Yandex announced the opening of access to the YandexGPT API for all users. Also, charging rules and quotas for requests began to operate, which can be increased by contacting technical support.

Access to the YandexGPT API is provided through the Yandex Cloud cloud platform. Previously, companies could take advantage of the technology by submitting an application to Yandex. According to Yandex's own data, during the testing of the YandexGPT API, by the beginning of December 2023, more than 20 thousand requests were received from banks, retailers, IT and other companies.

Thus, the online service for business "Bitrix24" introduced YandexGPT into the work of its AI assistant CoPilot, endowing it with such functions as recording a conversation with a buyer, filling out order details, writing messages, delivering a task and editing text.

Shoe chain Ralf Ringer uses the YandexGPT API to automatically answer customer questions on marketplaces and comment on product reviews.

According to the results of testing and the first pilots of companies with the YandexGPT API, we saw great interest and potential for using the technology. The business uses the neural network not just to automate individual business processes, but to develop IT products, "said Grigory Atrepyev, director of product development at Yandex Cloud.

As part of the development of the YandexGPT API, the Yandex Cloud Boost AI grant program was launched, whose participants can receive 1 million rubles to use the YandexGPT API to create chat bots and smart assistants, generate, structure and summarize text information in business applications and web services. Also, Yandex Cloud Boost API users will be provided with advice from architects on working with the next generation neural network and discounts on further use of the YandexGPT API.^[2]

Using YandexGPT to write ads on "Авто.ру"

On October 10, 2023, Yandex announced that all Авто.ру sellers will now be able to write an advertisement for the sale of a car using the YandexGPT generative neural network. This functionality allows car sellers to save time and energy - the neural network will take on the task of creating a description of the car for sale. Read more here.

Using the YandexGPT neural network in Yandex.Lavka

Yandex.Lavka"" began adding product information to its application using neuronets YandexGPT and technology (OCR Optical Character Recognition). They analyze the photo of the label of each product and independently fill out its card. This will allow the service to triple the speed of updating product information, simplify the work of content managers, and buyers will be able to quickly learn about new products on the platform. About this in "" told Yandex September 12, 2023. More. here

YandexGPT 2 Launch

On September 7, 2023, Yandex announced the launch of YandexGPT 2. This language model can solve more different types of problems than the first version of the neural network, and the quality of its answers has become more accurate.

As reported, YandexGPT 2 is able to adapt the response to different audiences and rewrite it in other styles and genres. For example, a neural network can explain a complex term to a child in simple words or retell a director's decree in the form of an understandable letter for internal distribution to employees. YandexGPT 2 is able to analyze any texts up to a thousand characters long that the user will add to the request, answer questions to them and structure information.

To evaluate the quality of YandexGPT 2, both versions of the model responded to the same set of requests. Then Yandex AI trainers compared the results taking into account many parameters, including accuracy, completeness, actual correctness, grammatical and stylistic correctness and others. YandexGPT 2 responds better in 67% of cases.

The proportion of requests of different types in which YandexGPT 2 responds better than the previous model:

• Text generation - 69%.
• Retelling and text analysis - 68%.
• Generation of ideas - 66%.
• Stylization for a specific audience or character - 62%.
• Actual requests - 62%.

To create YandexGPT 2, the developers increased the number of model parameters and optimized its learning process. They strengthened control over the quality of the source data, supplemented the data set with additional examples of requests and synthetic data - materials artificially generated based on real information.

YandexGPT 2 is already available in Alice's "Let's come up with" mode in the Yandex application, Yandex Stations, TVs with Alice, Yandex Browser, on the page with search results and on the ya.ru. In Search and Browser, chat with Alice can now be expanded to the full screen for more convenient work.

Yandex told how it will earn on its counterpart ChatGPT

In September 2023, Yandex"" launched YandexGPT 2. The new language model can solve more different types of problems than the first version, and the neuronets quality of its answers has improved significantly. Dmitry ChatGPT RBC Masyuk, director of the search and advertising technologies business group, spoke about how the company is going to make money on its analogue.

According to him, it is planned to introduce paid access for companies that connect to the YandexGPT software interface (API) to obtain services based on the neural network - creating a description of goods, retelling the main one of the texts, their stylistic improvement, increasing the productivity of the customer support service, etc. By September 7, 2023, Yandex received a request to connect to the YandexGPT API from about 6 thousand companies, at the moment access is provided free of charge, Masyuk said.

Yandex is also going to add new features to YandexGPT, which will be available to users who pay for the subscription. Details in Yandex are not disclosed, but among the options under consideration are the creation of specialized neural networks, for example, legal or medical, which can analyze the texts of this topic, the creation of a tool to help developers, etc.

Dmitry Masyuk also announced Yandex's plans to increase the attractiveness of search for users by providing answers in the form of a thesis retelling and text analysis; it will be possible to formulate more complex requests in the form of a coherent conversation. The increase in the number of search sessions should increase advertising revenues, according to the company.

Masyuk did not specify how much the company can earn through the monetization of YandexGPT. He only noted that Yandex is testing solutions based on YandexGPT in three of its own services and estimates the economic effect of their implementation at 1.2 billion rubles in 2024.^[3]

Use in the creation of an educational neural network for the study of informatics

Platform. Yandex The school prepares an educational one neuronet that will help children and adolescents learn. to programming It will allow schoolchildren of different ages to solve problems at to computer science a comfortable pace and will adapt to the level of knowledge of each student. This was Yandex announced on August 3, 2023. educational Yandex specialists and computer science teachers are working on creating the network. They will train a neural network based on YandexGPT (YaLM 2.0). More. here

Availability for testing in Yandex.Cloud

On June 19, 2023, the company Yandex.Cloud (Yandex.Cloud) announced that YandexGPT (YaLM 2.0) had become available for testings on the cloudy platform. Yandex Cloud More. here

Add to Masterpiece

On June 15, 2023, Yandex announced the addition of a large language model of the latest generation YandexGPT (YaLM 2.0) to the Masterpiece. Thanks to this, users can receive in response to a text request not only pictures, but also meaningful posts with a title and a suitable illustration. This opens up even more opportunities for creativity, fun and the preparation of useful content. Read more here.

Creating a neural network

On February 1, 2023, it became known that Yandex is developing its version of the ChatGPT generative network as part of the development of a language model from the YaLM family (Yet another Language Model, a neural network based on GPT-3 solutions from Open AI). The new project was named YaLM 2.0.

As a source familiar with Yandex's plans told Izvestia, by the end of 2023 the company is going to integrate this neural network into its services.

Yandex is indeed developing YaLM 2.0, its version of the ChatGPT generative network. The new neural network will become part of Search, Alice, Mail and other Yandex services. Yandex already partially uses generative neural networks in the formation of search results, but now they rather have an auxiliary role - they help to navigate, but do not generate coherent answers, "the company said.

They explained that the introduction of YaLM 2.0 will allow the search to generate answers itself using knowledge of the entire digitized world. "Alice" will become smarter, it will be better to hold the context of the conversation and will be able to answer more difficult questions. This could be a breakthrough in the quality of search engines and voice assistants, Yandex said.

The new neural network will allow developing and improving services addressed to domestic users, says Sergey Polovnikov, head of the Content-Review project.

To promote digital companies in a competitive market, it is necessary to constantly improve artificial intelligence technologies, which, in fact, is happening now, the expert added in a conversation with the newspaper (the article was published on February 1, 2023).

The new neural network is a very powerful technological step designed to show users and investors that, despite all the perturbations, Yandex remains a world-class advanced digital company, said Vladimir Zykov, director of the Association of Professional Users of Social Networks and Instant Messengers.^[4]

Notes