Sber AI

Sber AI is the artificial intelligence division of Sberbank, the Russian state-controlled financial-services company headquartered in Moscow. Sberbank, founded in 1841 as the Russian Empire's savings-bank monopoly and recapitalized as the modern banking institution in the Soviet and post-Soviet eras, holds approximately 50 percent of all Russian household deposits and is the country's dominant retail bank. Its majority shareholder is the Russian Ministry of Finance (50 percent plus one share). Under Herman Gref, the long-serving Chief Executive Officer (since 2007), Sberbank has expanded from a banking franchise into a broader technology conglomerate spanning food delivery (SberMarket), e-commerce, music streaming (Zvuk), media, mobility, and a substantial AI research and product line. Sber AI develops the GigaChat foundation-model family, the Kandinsky image- and video-generation model line, the Russian-language NLP suite, and AI capabilities embedded across the SberDevices smart-speaker line and the broader Sberbank consumer product surface. As of April 2026, Sber AI is one of the two principal Russian AI organizations alongside Yandex, operating in a domestic-market that has been substantially closed off from Western AI infrastructure since the February 2022 invasion of Ukraine and the subsequent Western sanctions packages that placed Sberbank on the US Specially Designated Nationals list.

At a glance

• Founded: Sber AI established as a distinct division in 2018 to 2019, with predecessor AI research at Sberbank dating to the early 2010s. Sberbank itself was founded in 1841.
• Status: AI division of Sberbank, the Russian state-controlled bank. The Russian Ministry of Finance holds 50 percent plus one share.
• Funding: Operates within Sberbank's commercial revenue base. Sberbank reported approximately RUB 1.6 trillion in 2024 revenue (approximately $18 billion). Sber's overall technology spend is reported in the multi-tens-of-billions of rubles annually.
• Senior leadership: Herman Gref, Chairman and Chief Executive Officer of Sberbank since November 2007. Former Russian Minister of Economic Development under Putin.
• Other notable leadership: Andrey Belevtsev, former Senior Vice President for Technology at Sberbank, oversaw the AI organization through 2023. Albert Yefimov, Vice President responsible for AI research direction. Senior research leadership across the GigaChat and Kandinsky programs.
• Open weights: Yes, partial. The Kandinsky image-generation models are released open-weights through the Sber-affiliated ai-forever Hugging Face organization. GigaChat models are closed.
• Flagship outputs: GigaChat foundation-model family (initial release April 2023; GigaChat Pro and Max successors); Kandinsky 2.x and 3.x image-generation models; Kandinsky Video (text-to-video); the Salute voice assistant on SberDevices smart speakers and SberBox set-top boxes.

Origins

Sber AI's lineage traces to Sberbank's transformation under Herman Gref, who took the Chief Executive Officer role in November 2007 with an explicit mandate to modernize what was then a Soviet-legacy bank. Through the 2010s, Sber built one of the larger technology and engineering organizations in Russia, with the AI investment ramping after 2017 as the broader Russian state articulated its national AI strategy. The May 2018 Sberbank announcement of Salute, an in-house voice assistant, signaled the company's positioning as a consumer-AI platform rather than a pure banking-technology operation.

The 2019 launch of the SberDevices subsidiary brought together the AI organization with the consumer-electronics product line (smart speakers, set-top boxes, smart displays) under unified leadership. The April 2023 release of GigaChat 1.0, a Russian-language foundation model, gave Sber a direct response to OpenAI's ChatGPT, which had been formally restricted from Russian users in early 2023. The Kandinsky line had launched earlier — Kandinsky 2.0 in November 2022, immediately after the Stable Diffusion release wave — and quickly attained substantial Hugging Face download volume from Russian-language users and the broader open-source community.

The February 2022 invasion of Ukraine restructured Sber AI's operating context. Sberbank was placed on the US Office of Foreign Assets Control Specially Designated Nationals list in April 2022, on the EU sanctions list, and on the UK sanctions list. The sanctions packages restricted Sberbank's access to Western financial infrastructure, foreign-exchange markets, and (in practical effect) to leading-edge Western AI hardware. Through 2022 and 2023, Sber retained access to Nvidia GPU inventory it had stockpiled before sanctions but lost the ability to procure new H100 and successor chips through normal channels. The company has subsequently reported expanded use of Chinese-fabricated AI accelerators and domestic Russian compute infrastructure.

The 2023 to 2026 period saw GigaChat iteration through GigaChat Pro (2024) and GigaChat Max (2025), with the company claiming Russian-language benchmark parity with GPT-4-class international models on domestic tests. Sber has not submitted GigaChat to international benchmarks like LMArena, and the closed-weights distribution prevents independent verification.

Mission and strategy

Sber AI's stated mission is to advance Russian-language AI capability and to embed AI across Sberbank's consumer and enterprise product surface. The strategy combines three threads. First, the GigaChat foundation-model line and an enterprise GigaChat API offering for Russian corporate customers locked out of Western API providers. Second, the Kandinsky image- and video-generation line, distributed open-weights to anchor Sber's positioning in the global open-source community and to attract domestic developer engagement. Third, deep integration of AI capabilities across the SberDevices product line, the SberBank mobile app, and the broader Sber consumer-services portfolio (including the Salute voice assistant).

The competitive premise is that the Russian-language and Russian-domestic-market segment of the global AI market is large enough to sustain a domestic foundation-model lab, particularly under conditions where Western API providers are formally restricted from operating in Russia. The premise is closely paralleled by Yandex's positioning, with both companies competing for the same domestic enterprise and consumer customers.

Models and products

• GigaChat. Russian-language foundation-model family. GigaChat 1.0 (April 2023), GigaChat Pro (mid-2024), GigaChat Max (2025). Closed-weights, available through API and through Sber consumer products. Reportedly trained on Russian-language web crawl, Sber-internal corpora, and licensed content.
• Kandinsky 2.x and 3.x. Open-weights image-generation models. Kandinsky 2.0 (November 2022), 2.1 (April 2023), 2.2 (July 2023), 3.0 (November 2023), 3.1 (April 2024). Strong Hugging Face download volume from Russian-language and broader open-source users.
• Kandinsky Video. Text-to-video model line, released open-weights. Smaller scale than the leading commercial video generators but the principal Russian-language video-generation option.
• Salute voice assistant. Embedded across SberDevices smart speakers (SberPortal, SberBoom), SberBox set-top boxes, and the Sberbank mobile app. Russian-language conversational interface for banking, smart-home, and content-discovery use cases.
• Enterprise GigaChat API. Russian commercial customers including banks, telecoms, and government agencies. Operates as the principal Russian-language alternative to OpenAI, Anthropic, and Google APIs.
• GigaCode. Code-completion model, integrated into Russian developer tooling.

Distribution channels combine API access for enterprise customers, embedded integration across SberDevices and Sber consumer products, and open-weights distribution through Hugging Face for Kandinsky.

Benchmarks and standing

Sber AI's evaluation framework is anchored on Russian-language benchmarks (Russian SuperGLUE, MERA, and Sber-internal evaluation suites) where the company has reported leading scores against domestic competitors. The closed GigaChat distribution and the lack of submissions to LMArena or other international benchmarks limit independent verification of the company's GPT-4-parity claims. International coverage has been generally skeptical of GigaChat reaching frontier-international parity but acknowledges that the Russian-language domestic-benchmark performance is competitive within the local market.

The Kandinsky line has stronger international visibility through the open-weights releases. Hugging Face download volumes have been comparable to second-tier international image-generation models, with Kandinsky 3.x considered a credible alternative to Stable Diffusion variants for users wanting an open-weights option outside the Stability AI lineage.

Leadership

As of April 2026, Sber AI's senior leadership includes:

• Herman Gref, Sberbank Chairman and Chief Executive Officer. Public face for the broader Sber strategy.
• Albert Yefimov, Vice President responsible for AI research direction.
• Senior research and engineering leadership across the GigaChat foundation-model program, the Kandinsky image-generation team, and SberDevices.

The senior leadership has been comparatively stable through the post-2022 period. Industry coverage has reported emigration of senior Russian engineers to international roles after 2022, but Sber's research-leadership ranks have publicly continued under domestic operation.

Funding and backers

Sber AI operates within Sberbank's commercial revenue base. Sberbank's 2024 revenue was approximately RUB 1.6 trillion. The Russian Ministry of Finance is the majority shareholder (50 percent plus one share); the remaining shares trade on the Moscow Exchange (SBER), with international trading suspended after the February 2022 sanctions packages. Sber has reported substantial domestic investment in AI infrastructure, including Russian-domiciled data centers and partnerships with domestic chip-design organizations.

Industry position

Sber AI occupies a domestic-market position rather than a global-frontier position. Within Russia, Sber and Yandex are the two organizations with the engineering scale and capital to sustain frontier-foundation-model research, and the two compete for enterprise customers, developer mindshare, and government-aligned strategic projects. Internationally, Sber's profile is anchored on the open-weights Kandinsky line; the closed GigaChat is functionally invisible to non-Russian-language users.

The structural risks are political and supply-chain. Sanctions persist as of April 2026 with no public sign of relaxation. Access to leading-edge AI accelerators continues to be constrained, and Sber's domestic compute infrastructure relies on a combination of stockpiled Western chips, Chinese-fabricated alternatives, and domestic Russian production that has not approached frontier-international capability. Continued sanctions extension and any escalation tied to the Ukraine conflict would deepen these constraints.

Competitive landscape

• Yandex. Direct domestic peer. The other principal Russian AI organization. YandexGPT and YandexART are the closest analogs to GigaChat and Kandinsky. Yandex completed its July 2024 corporate restructuring (separating Russian and international operations into two companies) which reframed Yandex's positioning relative to international markets.
• Western frontier API providers (OpenAI, Anthropic, Google DeepMind, Mistral AI). Formally restricted from operating in Russia. The market gap is what GigaChat exploits domestically.
• DeepSeek, Z.AI, Moonshot AI, Alibaba Qwen / DAMO. Chinese open-weights models that are usable in Russia and that some Russian developers have substituted for GigaChat for cost or quality reasons.
• Stability AI, Black Forest Labs. Western open-weights image-generation peers; Kandinsky competes with these for the international open-source-developer audience.
• Domestic Russian AI startups. Smaller Russian AI companies including MTS AI, Tinkoff Research, and adjacent Russian banks' AI divisions, all operating at materially smaller scale than Sber and Yandex.

Outlook

• The continued GigaChat iteration cadence and any independent verification of frontier-international parity claims.
• The Kandinsky open-weights line's continued Hugging Face traction.
• The compute-supply trajectory under continued Western sanctions.
• Any change in the Western sanctions regime tied to Ukraine-conflict resolution.
• The competitive dynamic with Yandex for Russian enterprise customers.
• Senior researcher emigration trends and the resulting effect on Sber's research output.

Sources

• Sber AI. Division reference.
• GigaChat. Foundation-model product.
• Kandinsky on Hugging Face. Open-weights image-generation distribution.
• Sberbank. Parent-company reference.
• Herman Gref Wikipedia. Sberbank CEO reference.
• US OFAC Specially Designated Nationals list. Sanctions reference.