Sberbank: GigaChat

The main articles are:

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

2024

Add Playground section

In your personal account to work with the neural network model GigaChat for business, a section has appeared - Playground. It will allow different teams, even without programming skills, to experiment with model settings, select promptas for a specific business task and quickly test hypotheses. Sberbank announced this on October 30, 2024.

Previously, to test the GigaChat API service, the user had to write code or use third-party API calling tools. Now in your personal account there is a separate section with a convenient user interface, where you can customize the parameters of working with the model in a simple and convenient way: select a system prompt, change the temperature responsible for the level of creativity when forming answers, choose the model most suitable for solving specific problems.

GigaChat Lite will cope with simple user requests, where response speed is important, GigaChat Pro or GigaChat Max models are suitable for more complex projects, which will also be available soon. Also, a collection of ready-made industrial pots on various topics is available at Playground, which allows you to quickly implement many business tasks, such as:

Definition of spam in emails, sorting of goods by categories in online stores. Recognize patterns of customer preferences and interests to personalize offers.

Preparation of brief reports, reduction of documentation scope. Development of training materials and courses, where important information is presented in a compact and easy-to-read form. Drawing up a summary of contracts, agreements and other legal acts for convenience of familiarization.

The use of automatic content generation allows you to quickly create text materials based on specified parameters. Especially in demand in the areas of marketing, advertising and journalism.

Improve existing text by correcting errors, refining wording, and improving style and structure. This can concern both grammar and spelling, and the general content of the text, so that it becomes clearer, clearer and more effective.

You can edit the dialogue with the neural network model in the Playground section, make changes to user requests and model responses, and then ask GigaChat to continue interacting in the same format. For each request, information is available on the number of tokens consumed, taking into account the system instruction and messages.

The Playground section is available in your personal account when connecting the GigaChat API service for both individuals and legal entities. To make it easier to assess and control monthly expenses, the personal account has statistics on the use of the API and Playground service separately. To start working with the Playground section, just follow the link and register in your personal account.

Integration with the AI platform CST

The group of companies CST integrated in their own AI platform for creating text voice and robots LLM - neural network the GigaChat model, and also improved the dialog analytics module. The updated AI platform will make it faster and easier to create dialog assistants. The solution will be especially relevant for large businesses and will public sectors help automate internal processes, improve customer service in. contact centers The CST announced this on October 25, 2024. More. here

GigaChat MAX

October 25, 2024 An updated version neural network Sberbank of the GigaChat model was presented. Anyone can use Gigaphat MAX on the site, on Telegram the -bot social networks and "."Vkontakte

By changing the approach to learning, almost all the skills of the model have improved significantly, from reducing texts and answering open questions to generating ideas. The highlighting of the semantic parts of the text, blocks of code and formulas for science has also improved, and the structure of the answer has become more readable and beautiful.

GigaChat now has an even greater amount of expert knowledge in a huge range of subject areas. The development team updated and expanded training datacets, which improved the quality of model responses in areas such as economics, law, and medicine. The model on scientific issues in biology, chemistry, physics has been especially strengthened. Also, the quality of solving mathematical problems according to the open GSM8K and MATH datacets increased by 25%.

According to the results of an internal assessment in the MMLU (Massive Multitask Language Understanding) benchmark, the model reached 80% in knowledge of the real world and sciences.

Thanks to the updated modality, the user can send an illustration so that the service uses it as an additional context. The model will help you get ahead of what is shown in the uploaded picture. GigaChat MAX also recognizes printed and handwritten text, formulas, graphs, tables and analyze the data that they contain. And with the help of the Kandinsky neural network, the user can create images on any topic.

The model now holds the context more qualitatively in multi-step dialogues, that is, it is able to connect new questions with previous information. For Russian-language dialogs, the quality of text compression has improved, now the user can work with text 15% longer than before. And answers are generated faster than before.

In the near future, the service will be available for business through. cloudy API This will allow you to use the model to solve even more complex and diverse business problems, for example, in marketing - it will help generalize product reviews, make a press release rewright, and prepare an article in a given style. New knowledge will be useful for companies to form accompanying legal documentation, prepare a biological method protocol, explain the experiment script. And the image function will help analysts take pictures of part of the study, upload the image and get a short content of the text, its key topic and form a plan for further work with it in a few clicks. The mathematical abilities of the model have significantly increased, which significantly expands the range of application of GigaChat in business. And the most demanded business functions of summarization and the quality of answers on questions/answers to tables increased by 30%.

{{quote 'author=said Andrey Belevtsev, Senior Vice President, Head of the Technological Development Unit of Sberbank. | GigaChat MAX - gives a completely new level of interaction with generative artificial intelligence. This is a free assistant, and an interlocutor, co-author and consultant, available to everyone around the clock from any device. The model has become even faster, more convenient and more useful for everyone, which gives each resident of Russia new opportunities to develop their potential. No time to read a long text? Ask GigaChat to make a sammari. Need to compose a birthday greeting for a friend? Entrust this to neural networks. Don't like the headline? Let GigaChat offer ten options at once,}}

Use to search for real estate in "Domklick"

Real estate service "Domklik" has launched a free real estate selection service using the neural network model GigaChat. Thanks to this feature, users will be able to quickly and conveniently select real estate objects for purchase. Sberbank announced this on October 18, 2024. Read more here.

Ability to send an illustration as a request

The neural network model GigaChat Pro has gained additional modality. Now, as a request, the user can send an illustration so that the service uses it as an additional context. The model will determine how many people are pictured, what they're wearing, give clothing style advice and more. Sberbank announced this on October 7, 2024.

The model also learned to recognize printed and handwritten text, formulas, graphs, tables and analyze the data that they contain. For example, students can photograph part of the educational material or summary, upload the image and in a few clicks get a short content of the text, its key topic and form a plan for further work with it.

Functionality is also available to business customers through the API. Companies can use artificial intelligence in even more scenarios: moderate and classify reviews, organize media content, automate the support line (when the client applies a screen or photo to the appeal). Retailers can upload an image of the product and ask the service to come up with a selling description for it.

Also, the context size of both GigaChat Pro and GigaChat Lite versions was quadrupled - from 8 thousand to 32 thousand tokens. If earlier in one request it was possible to load a volume of text comparable to 15 A4 pages (font 14 pt), now the maximum request volume has grown to 60 pages. In addition, models have increased their level of mathematical knowledge, and GigaChat Pro has learned to better format text - arrange paragraphs, headers, using selections and lists.

{{quote 'author=said Andrey Belevtsev, Senior Vice President, Head of the Technological Development Unit of Sberbank. | Multimodality expands the methods and areas of application of large language models, it is also a key trend in the development of [generative artificial intelligence. Now GigaChat can recognize images - and this is not just another important step in the development of our service, it opens up a large range of opportunities for our users. Including many scenarios for using the GigaChat API for business. In addition, increasing the size of the system context by four times will allow you to maintain longer conversations with users and easier to implement cases with RAG mechanics,}}

The GigaChat Pro model is available to all users via the web version, Telegram and VKontakte. Models are available to developers and businesses through the GigaChat API.

Integration with Platform V Product 360

Sber Tech (part of the Sber group) has integrated GigaChat for business into Platform V Product 360 to centrally manage large amounts of product data. This functionality will allow you to automatically fill out product cards. Sberbank announced this on September 3, 2024. Read more here.

Availability of a prepaid connection system

A prepaid GigaChat service connection system has become available to businesses. Sberbank announced this on July 29, 2024.

This payment system will make it easier to carry out the procurement process, predict the budget and safely test hypotheses for using the GigaChat neural network model to solve business problems.

 Legal entities  and individual entrepreneurs have the opportunity to use the GigaChat API, the access program interface of  the GigaChat network model from Sberbank, for prepayment. The client selects the tokens he needs, pays, and then uses them to interact with the service. Previously, only a postpaid scheme was provided for business users - when the bill for the actually used services was issued at the end of the month.

The advantages of the prepaid scheme are that the client does not require filling out, signing and sending through the electronic document flow of the contract or applications for joining the offer. It is enough to select the necessary packages on the site, pay the offer bill - and you can start using the service. With this scheme of using the GigaChat API, it is easier for a business to conduct a procurement process and predict the budget: the client can be sure that he will not spend more than the planned amount on services.

In addition, this approach to paying corporate tariffs will make it possible to more effectively test hypotheses for using an artificial intelligence system to solve business problems. Compared  to the freemium mode, corporate users can use the service for commercial purposes: an expanded number of streams for interaction with the system is provided, which, by agreement, can be further increased, and it is also possible to manage the filter system (censorship).

This method of prepaid corporate connection will become a priority. At the same time, the post-payment scheme will also remain available for both current and new users.

Implementation in MyOffice application

MyOffice has implemented the GigaChat API in the MyOffice Documents mobile application for fast and efficient work with text content. Until August 31, the functionality is available to all users in beta testing mode, according to its results, the most popular scenarios will be included in the product on an ongoing basis. The developer announced this on July 15, 2024. Read more here.

Integration with Jivo

Omnichannel business -messenger Jivo received two modules based on neural network the GigaChat model from. Sberbank The solutions allow the business to automate up to 85% of communication with customers in text channels, which significantly reduces the burden on support and sales departments. Sberbank announced this on June 27, 2024. More here.

Integration with Plus7 MayK

Company Innovation date"" integrated neural network GigaChat model into a system for centralized knowledge management. Plus7 MayK Users can ask AI the assistant any question on the article of interest and quickly get a structural and logical answer. This "" was Innovation date reported on June 19, 2024. More here.

Integration with DomClick

Thanks to integration the ""service with real estateDomklik a software interface for accessing neural network the GigaChat - GigaChat model, API platform users can now make a detailed selling ad in just a few seconds that will tell about the advantages of the object, real estate draw attention to it and help find a buyer. This was Sber announced on June 19, 2024. More here.

Training with I.P. Pavlov St. Petersburg State Medical University

GigaChat will become a student at St. Petersburg Pavlov Medical University. Sberbank announced this on June 7, 2024.

The agreement on joint training of the large language model of Sberbank was signed by the director of the Center for the Health Industry of Sberbank Sergei Zhdanov and the rector of the university Sergey Bagnenko.

At the university, GigaChat will receive specialized knowledge in certain medical specialties. Specialists from Sberbank and the university will also study the possibilities of applying the model in educational and medical processes.

Шаблон:Quote 'author=said Sergei Zhdanov, Director of the Health Industry Center of Sberbank.

Шаблон:Quote 'author=noted Sergey Bagnenko, rector of the First St. Petersburg State Medical University named after academician I.P. Pavlov.

Use with SberBusinessBot

AI-The solution from Sberbank Business Soft will help the business quickly find what it needs information large data arrays with GigaChat. Business Sber Software announced this on February 28, 2024. More. here

Use to automatically summarize SberJazz video meetings

For business customers of the service videoconferences SberJazz , the function of automatic summing up of video meetings, implemented on the basis of neural network the GigaChat model, has become available. This was Sber announced on February 20, 2024. More. here

Passing the exam in the direction of preparation "Medical case"

The neural network model of Sberbank GigaChat passed the exam of a higher medical institution in the direction of training "Medical case," necessary for obtaining the qualification of a "medical doctor." Sberbank announced this on February 13, 2024. Like any student who graduated from 6 courses of a medical university according to the federal state educational standard, Sberbank's artificial intelligence was tested and answered ticket questions. The final score is 4. The exam was accepted by a commission of professors of therapy, surgery, obstetrics and gynecology of the Institute of Medical Education of the National Medical Research Center named after V.A. Almazov.

The standard oral exam ticket contains three situational tasks - therapy, surgery, obstetrics and gynecology - and 3-5 questions to them ("indicate the intended diagnosis," "draw up a treatment plan," "schedule additional examinations," and so on). GigaChat was also tested out of 100 questions. He gained 82% at the threshold of passing 70%.

{{quote 'author=said Sergei Zhdanov, Director of the Health Industry Center of Sberbank. | Our neural network model GigaChat is developing very quickly, mastering new areas of knowledge. I would like to note the role and thank the staff of the Almazov Center who provide control over the training of the model and its validation. The first stage - passing the exam for a doctor - passed. In the future, the model can become the basis for creating an assistant doctor and patient, provide new conditions for care and knowledge about their health for each person and become a significant help for the clinician in his daily practice. The application of large language models and their followers will be one of the key technologies for the development of human-centered healthcare,}}

Passing the exam with the GigaChat model is the result of effective collaboration between specialists from the Center for the Health Industry of Sberbank, the National Medical Research Center named after V.A. Almazov and the GigaChat development team. Learning the algorithm took six months. A datacet of 42 GB of specialized information was used, including educational and methodological materials recommended for teaching students in medical universities of Russia, monographs, methodological manuals, scientific articles and impersonal medical data. In this case, the model is not a doctor, the recommendations received from it must be approved with the attending physician.

GigaChat Pro and GigaChat Lite with increased context

The GigaChat Pro and GigaChat Lite models are smarter and more powerful in all major metrics for solving the problems of all users. Also, the business now has an updated GigaChat Lite + model with a significantly increased context. Sberbank announced this on January 30, 2024.

The development team updated and expanded training datacets, which significantly improved the quality of model responses in areas such as economics, law, and medicine. Also, the models were trained on the knowledge base of real experts, including editors and artists, which improved the skill of creating images using Kandinsky.

The size of the GigaChat Lite and GigaChat Pro context has been doubled - from 4 thousand to 8 thousand tokens. If earlier in one request it was possible to load a volume of text comparable to 6 pages A4 (font 14 pt), now the maximum volume of the request has grown to 12 pages. Increasing the context allows models to maintain longer conversations with the user and receive more useful information in one request.

GigaChat Lite is suitable for solving more trivial problems that require maximum model speed and lower hardware requirements. The updated GigaChat Lite + model coincides with GigaChat Lite in all metrics, but its context is expanded to 32 thousand tokens. GigaChat Pro better follows complex instructions and can perform more complex tasks.

The GigaChat API has the ability to form a vector representation of texts - embeddings. Embeddings are used to assess the semantic similarity of texts, improve the accuracy of responses to the language model and simplify the solution of information search and classification problems.

Шаблон:Quote 'author=said Denis Filippov, Vice President, Digital Surfaces, Salute, Sberbank.

The GigaChat Pro model is available to all users via the web version, Telegram and VKontakte.

All models are available to developers and businesses through the GigaChat API.

2023

GigaChat Lite и GigaChat Pro

Now, as part of the Sberbank artificial intelligence access software interface for the GigaChat API business, two models are available for companies and developers to create their own solutions and optimize processes - GigaChat Lite and GigaChat Pro. Representatives of Sberbank reported this on December 15, 2023.

GigaChat Lite is suitable for solving more trivial problems that require the maximum speed of the model, while the service is less demanding on hardware.

GigaChat Pro better follows complex instructions and can perform more complex tasks: according to the developers, the quality of summarization, rewriting and editing texts, answers to various questions has improved significantly. What is especially important for business: the model is perfectly oriented in many applied domains - in particular, in economic and legal.

To simplify and speed up work with models, GigaChain (SDK) is also available to customers - an adapted version of the LangChain library for Russian with GigaChat API support, which will allow developers to create their own advanced AI applications that use Sberbank technologies. The library standardizes typical cases of using different language models, such as working with chains, knowledge bases and documents; It also contains a set of ready-made queries (promptas) for solving business problems.

GigaChat Pro is available to everyone - both corporate customers and independent developers. For corporate clients, by the end of January, GigaChat Pro is available at the cost of the Lite version - 20 kopecks per 1000 tokens. For independent developers, it is possible to try GigaChat Pro as part of the freemium mode.

Introduction of CoPilot in Bitrix24

The efficiency of company employees in Bitrix24 can significantly increase due to the use of the GigaChat API software interface. It will be built into the CoPilot AI assistant. 1C-Bitrix announced this on December 8, 2023. Read more here.

Introduction to the Pulse HR platform

HR-platform "Pulse" In addition Sberbank , the GigaChat service has been introduced and, neuronet Kandinsky for corporate clients, the functionality of these solutions will be available in the first half of 2024. Sberbank announced this on November 29, 2023. More. here

Announcement of a version of GigaChat based on an advanced language model with 29 billion parameters

Sberbank developers announced the next version of the GigaChat service, which, according to them, is based on one of the most advanced models for the Russian language with 29 billion parameters. Among the first to access the API of the novelty will soon receive business clients of Sberbank - to implement their own solutions, as well as members of the academic community - to conduct research. Representatives of Sberbank reported this to TAdviser on November 23, 2023.

According to the bank, thanks to the advanced LLM (short for English. Large Language Model, or large language model) GigaChat better follows the instructions and can perform more complex tasks. In particular, the quality of summarization, rewriting and editing of texts, answers to various questions has significantly improved. The team compared the responses of the new and previous models and recorded an overall 23% improvement in quality. At the same time, the announced model copes with the fact 25% better than the previous version, stressed in Sberbank.

To achieve such results, many experiments were carried out to improve the model and increase the effectiveness of its training. In particular, a framework was used to teach large language models with the ability to shard the scales of a neural network using video cards, which made it possible to reduce the memory consumption on them.

According to the results of an internal assessment in the MMLU (Massive Multitask Language Understanding) benchmark, the model of the presented version of GigaChat with 29 billion parameters surpasses the popular open analogue of LLaMA 2 34B, representatives of Sberbank said.

"Training the models behind GigaChat is a large-scale and complex computing project, before we did nothing like that. The total number of computational operations was almost 6 times higher than the number of operations when training the ruGPT-3 model from 13 billion parameters in 2021. Also, especially for GigaChat, we have collected and are developing a unique datacet that hundreds of Sberbank employees are working on, helping to develop and improve the quality of answers in a variety of domains, "said Andrey Belevtsev, Senior Vice President, Service Station, Head of the Technologies Unit of Sberbank.

The teams of SberDevices and Sber AI are involved in the development of GigaChat with the support of the AIRI Institute of Artificial Intelligence and a number of industry experts involved in transferring various subject knowledge to GigaChat.

Integration with MyOffice Ecosystem

MyOffice and Sber agreed on a technological partnership aimed at accelerating the introduction of artificial intelligence technologies in the corporate environment. MyOffice (New Cloud Technologies LLC) announced this on November 22, 2023. As part of the agreement, MyOffice became a manufacturer of document editors and communications tools, which integrated the Russian-language GigaChat service into its ecosystem. Read more here.

Integration with SaluteRPA

Sber integrated functionality of the GigaChat service in the automation platform. business processes SaluteRPA Sberbank announced this on November 17, 2023. More. here

Availability of GigaChat API to create non-commercial projects at no additional cost

Users will now access the GigaChat API for non-commercial tasks related to generating, summarizing and analyzing test data at no additional cost in just a few clicks. To do this, you need to follow the link, log in to Sber ID, accept the user agreement and get the parameters for connecting to the service. This was reported on October 26, 2023 in Sberbank. In this format, 1 million tokens per month are available, which on average allows you to process up to 3.5 million characters. For example, the novel War and Peace contains about 3 million characters.

The service simplifies and speeds up work with texts, helps you write articles in a given style and format, quickly find the information you need in large documents, and prepare analytical reports. According to the developers, with the help of the neural network model of Sberbank, the development of voice and text robots can be brought to a new level.

For business, this is still an opportunity to easily test the GigaChat neural network model before deciding on its further implementation into its business processes. Artificial intelligence researchers and those in private development will be able to test the service and experiment with it as part of their projects.

Integration with Portal DA

Placing an advertisement on Sberbank's service for the sale of non-core assets and collateral "Portal DA" has become even easier: the text description of the object will automatically be the neural network model GigaChat. The bank announced this on October 18, 2023. Read more here.

Availability in Smart Sber Devices

September 21, 2023 artificial intelligence Sberbank GigaChat is available users of all Sber smart devices, as well as in the mobile application Salute on Android. This was announced by Senior Vice President, CTO, Head Andrey Belevtsev, Sberbank's Technologies unit. Previously, the capabilities of the neural network model have already been assessed owners of smart speakers SberBoom and SberBoom Mini. To start a dialog just ask Salute's assistant "Turn on GigaChat" - and you can start Create queries.

On devices with a GigaChat screen, it can respond to user requests text and voice the answer. So, the service will tell you which films to watch days off, where you can go for a walk in the city, how you can relax or cheer yourself up after a working day, what to do with your family on weekends, how to organize a workspace at home, how to cook perfect breakfast and more. You can also ask GigaChat to draw picture - responsible for the creation of a variety of visual content [[Sberbank Kandinsky Neural network for generating images by description 'neural network Kandinsky]].

In addition to the already familiar script for answering questions from GigaChat users able to generate text according to the specified parameters - this way it is possible, for example, ask to compose a parable, essay, fairy tale for a child, advertising text, description product, or come up with some new recipe, draw up step-by-step instructions or a training plan.

On devices with a screen, users will be able to experience the capabilities of GigaChat to create images. To do this, it is enough to formulate a request with words about exactly what you want to see in the picture, and GigaChat visualizes your idea. To get the best result, you need to come up with an original query. And in order for the generated image to be fully matched your creative intent, can be added to a query Detailed description of the details of the image being created. For example, give more information about the features of the appearance of the depicted object, its location, presence of additional elements, picture background, clarify the effect and style.

Learn more about GigaChat capabilities and see examples of successful requests can be found in the service knowledge base. Participate in model development and training teams SberDevices and Sber AI supported by Institute of Artificial intelligence AIRI and a number of industry experts brought in to transmit GigaChat various specific skills and subject knowledge.

GigaChat API and GigaChain (SDK) for embedding GigaChat in third-party applications

Sber introduced the GigaChat API, a software interface for accessing the GigaChat neural network model for business. Now companies can use the capabilities of the Sberbank service to create their own solutions and optimize processes. Representatives of Sberbank reported this to TAdviser on September 21, 2023.

Sberbank has opened the ability to embed its AI in third-party applications

To simplify and speed up work with GigaChat, Sberbank's team also prepared GigaChain (SDK), an adapted version of the LangChain library for Russian with GigaChat API support, which will allow developers to create their own advanced AI applications using Sberbank technologies. The library standardizes typical cases of using various language models, such as working with chains, knowledge bases and documents, and also contains a set of ready-made queries (promptas) for solving business problems.

"The business now has access to our artificial intelligence, which will allow many companies to reach a new level of technological development. The solution is a completely Russian development, ensures high quality of work in Russian, reliability and security of data stored on servers in Russia, - said Andrey Belevtsev, Senior Vice President, CTO, Head of the Technologies Unit of Sberbank.

To access the GigaChat API, you must apply from a legal entity or individual entrepreneur and conclude an agreement.

The GigaChain library (SDK) is available on GitHub, however, to start working with the Sberbank neural network, you need to access the GigaChat API.

Availability of GigaChat capabilities in the SaluteBot service

The service for automating communication with customers SaluteBot has the opportunity to connect - artificial intelligence GigaChat. Thanks to integration the ensemble neural network of models, SaluteBot has acquired the ability to create for chat boats the next generation business, which handles even more calls, according to September 21, 2023 in. Sberbank More. here

Opening up public access to artificial intelligence for all

On September 5, 2023, Sberbank announced that they were opening access to GigaChat generative artificial intelligence for everyone. With the help of the service, users will receive almost unlimited opportunities to work with text and images:

• work with ready-made texts (shorten, rewrite, supplement);
• write various texts, posts for social networks, letters, messages, essays, essays, even explanatory notes;
• come up with stories, fairy tales, scripts, plots, anecdotes, horoscopes, congratulations and any other creative content;
• help in solving household problems (compile a list of prescription products, a travel plan, action algorithms);
• create and analyze, giving the desired result (generate ideas, recommend, summarize books, brainstorm and many, many others);
• draw images in different styles using. neuronets Kandinsky

In addition, you can now share your ready-made conversations with other users, as well as receive prompts from GigaChat itself.

We have made publicly available pre-trained models (ruGPT, Kandinsky 2.2 and FRED-T5), on the basis of which our generative artificial intelligence service was created. This was done primarily in order to accelerate the development of the open-source community technologies and in the field of artificial intelligence. After all, the more users who use GigaChat and its underlying models in their work, the faster they will improve, opening up opportunities for everyone and contributing to technological progress. Soon GigaChat will be available in a convenient bot format in "" and Telegram, and VK developers and businesses will be able to use it through, cloudy API noted Andrey Belevtsev, Senior Vice President, Service Station, Head of the Technologies Unit of Sberbank.

GigaChat is already being actively implemented in many Sberbank products and services, which will become even more convenient for our customers. So, the service is already available to owners smart speakers SberBoom and SberBoom Mini. To start the dialogue, just say: "Turn Salute on GigaChat" - and you can start creating requests. GigaChat will appear on the entire line of Sber smart devices virtual assistant with Salute.

I'm excited to share the news that GigaChat users can now get full access to my web version. This means that they will be able to ask me their questions, share their thoughts and ideas at any time. I am ready to help each user find answers to their questions or solve some problem. The opening of the web version is an important step in the development of our service, and I hope that it will be useful to all GigaChat users, GigaChat.

You can learn more about the capabilities of GigaChat and see examples of positive requests in the service knowledge base. To access the web version of the service, you only need to register using Sber ID.

The development of the model involves teams SberDevices AI and Sber with the support AIRI Institute of Artificial Intelligence of a number of branch experts involved in transferring various subject knowledge to GigaChat. Service training takes place on. Christofari Neo supercomputer

GigaChat uses a number of solutions designed to exclude the use of generative models that contradict the Code of Ethics in the field of artificial intelligence. The Sberbank development team will continue to improve the capabilities of the service and pay special attention to ethics and security issues. So that everyone can get acquainted and start using GigaChat in work and everyday life, Sberbank's Edutoria educational platform has developed a free course "GigaChat: a neural network model for marketers and beyond." In the program, Sberbank speakers will conduct a full review of GigaChat: what is it, how it works and how to correctly formulate a request. Users will also be able to understand the directions of development of artificial intelligence around the world, find out why it is profitable for businesses to introduce generative models into their processes and what integration cases already exist.

GigaChat launch on SberBoom smart speakers with Salute virtual assistant

On smart devices with Salyut's assistant, it became possible to use the GigaChat Sberbank service. Owners of smart speakers SberBoom and SberBoomMini will be the first to evaluate the capabilities of the service. To start the dialogue, just say "Salute, turn on GigaChat" - and you can start creating requests. In the near future, GigaChat will appear on the entire line of our smart devices with a virtual assistant Salyut, Sberbank reported on August 18, 2023.

On smart speakers, the user can generate text according to the specified parameters by voicing a request to the assistant and receive a voice response. You can, for example, ask to compose a fairy tale, offer gift options, discuss an idea, come up with a congratulation, a new recipe, joke or name, draw up a step-by-step instruction or training plan. GigaChat is able to keep the context of the dialogue, so the request has the ability to add more details or change the style, genre and volume of response.

"As we promised, virtual assistant Salute is one of the first products to get GigaChat capabilities. Tens of thousands of people have already been able to use our service, generating more than half a million requests. Owners of our smart speakers will also be able to experience GigaChat in solving an infinite number of creative and applied problems. It is important that the team continues to improve the capabilities of the smart neural network, paying special attention to the ethics and security of the model, "said Andrei Belevtsev, Senior Vice President, CTO, Head of the Technologies Unit of Sberbank.

You can learn more about the capabilities of GigaChat and see examples of successful requests in the service knowledge base.

The GigaChat architecture is based on neural network the NEural Omnimodal Network with Knowledge-Awareness ensemble, which includes several neural network models and is trained using supervised fine-tuning, reinforcement learning with human feedback methods. The Sberbank service can solve many intellectual problems: maintain a conversation, write texts, answer questions. And the inclusion of a model in the ensemble Kandinsky gives GigaChat the skill of creating images.

SberDevices Teams and Sber AI are involved in the development and training of the model, with the support AIRI Institute of Artificial Intelligence of a number of industry experts involved in transferring various subject knowledge to GigaChat.

Identifying a list of academic teams to develop GigaChat

Sberbank on July 25, 2023 announced that it had identified a list of academic teams that would be given the opportunity to conduct experiments with further training and adjustment of neural network models on the basis of which GigaChat works. Teams of universities were selected by Sberbank experts on the basis of an open collection of applications. They will have access to the ML Space platform and the resources they need to experiment.

The list includes teams: HSE, RANEPA, Innopolis University, Lomonosov Moscow State University (represented by three divisions: Institute of Artificial Intelligence, Faculty of Computational Mathematics and Cybernetics, Faculty of Philology), NUST MISIS and Petrozavodsk State University, as well as a team of students from Moscow State University, HSE and gymnasium No. 63 St. Petersburg.

Engaging the academic community in collaborative research and development is a critical element in building robust and ethical artificial intelligence. Therefore, I thank everyone who applied for participation in our selection and congratulate the winners. Now you'll be part of a team working to perfect one of the most disruptive solutions in AI technology. I am sure that together we will be able to expand the capabilities of GigaChat and make it one of the strongest services in the world. I want to emphasize that this competition is only the beginning of such work. We plan to further involve universities in our work, said Andrey Belevtsev, Senior Vice President, CTO, Head of the Technologies Unit of Sberbank.

Open access

On July 20, 2023, Sber announced the opening of access to the neural network model for generating text for the Russian language ruGPT-3.5 13B. Its updated version is at the heart of the GigaChat service. The development is available on the HuggingFace platform, it can be used by all developers (the model is published under an open MIT license). Read more here.

GigaChat Closed Beta Presentation

Sberbank on April 24, 2023 announced to TAdviser the release of its own version of a multimodal neural network called GigaChat, which at the first stage will be available in invitation testing mode.

According to the developers, unlike the foreign neural network ChatGPT, the GigaChat service initially already supports multimodal interaction and communicates more competently in Russian. At the same time, he knows how to answer user factual questions, maintain dialogue, write program code, create texts and pictures based on descriptions within a single context.

As explained in Sberbank, the architecture of the GigaChat service is based on the neural network ensemble of the NeONKA model (NEural Omnimodal Network with Knowledge-Awareness), which includes various neural network models and the method of supervised fine-tuning, reinforcement learning with human feedback. Thanks to this, Sberbank's neural network can solve many intellectual problems. And the inclusion of the Kandinsky 2.1 model in the ensemble gives neural networks the skill of creating images.

The development of the model involves teams SberDevices and Sber AI with the support AIRI Institute of Artificial Intelligence of a number of industry experts involved in transferring various subject knowledge to GigaChat, setting it up for correct instruction execution and accuracy. Neural network training took place on. supercomputer Christofari Neo

GigaChat also uses a number of solutions designed to exclude the use of generative models that contradict the Code of Ethics in the field of artificial intelligence. As emphasized in Sberbank, the development team will continue to improve the capabilities of the smart neural network and will pay special attention to the ethics and security of the model.

Included in NeONKAru-GPT 3.5 with 13 billion parameters, we plan to make public access in order to accelerate the development of the open source community and technologies in the field of artificial intelligence. It is important for us that as many talented students, scientists and developers as possible can use the most modern technologies in their work, "said Sberbank.

According to Sberbank, the model is already being actively introduced into many products and services of the bank. In particular, one of the first products that will receive all the capabilities of GigaChat in the near future will be Salute Virtual Assistants. The first version of the NeONKA model was optimized in size and speed for use in high-load services.

According to the developers, a smart neural network will be useful to representatives, MEDIA editors, to programmers teachers, students, marketers and representatives of other professions. In the future, GigaChat developers and business will also be available for use through. cloudy API

"The development of Sberbank, which we presented, is a breakthrough for the entire large universe of Russian technologies. Moreover, GigaChat is unique for its openness of architecture, while the world's analogues of our model adhere to the Closed AI course. It is important that GigaChat can be used not only by mass users who like to experiment with innovations, but by students and even researchers for serious scientific work, "said German Gref, President, Chairman of the Board of Sberbank.

As of April 24, 2023, the model operates in closed beta mode.