GigaChat: Inside Russia’s Largest Bank-Built AI Model

Russia’s largest bank builds AI models most of the world has never heard of. GigaChat just got its biggest update yet — a 432-billion-parameter model released weeks ago. Here’s what it actually is, how it compares, and why a bank is the one building it.

What GigaChat Actually Is

GigaChat is a generative AI chatbot built by Sberbank — Russia’s largest bank, and increasingly a technology conglomerate in its own right. It launched in closed testing in April 2023, positioned explicitly as a Russian alternative to ChatGPT. By February 2024 it had passed 2.5 million users. It’s multimodal — it handles text, writes code, and generates images through an integrated tool called Kandinsky — and Sberbank has repeatedly emphasised that it’s built specifically to handle Russian language, slang, and cultural context better than foreign models trained primarily on English data.

The underlying architecture has a name most people outside Russia have never encountered: NeONKA — a neural network ensemble combining supervised fine-tuning with reinforcement learning from human feedback, developed jointly by Sberbank’s internal SberDevices and Sber AI divisions.

Coverage of GigaChat’s original 2023 launch. Independent production, not affiliated with Wangdoo.

What Just Shipped — GigaChat 3.5 Ultra

On July 6, 2026, Sber released GigaChat 3.5 Ultra, its largest and newest flagship model — a mixture-of-experts architecture with 432 billion total parameters, built on what Sber describes as a proprietary domestic architecture using linear attention technology. The company claims the model generates long text up to four times faster than its predecessor while using fewer computational resources and being nearly half as compact — and says it improves specifically on coding, mathematics, long-document handling, and autonomous agent tasks.

Sber made two notably open moves with this release. It’s free to use through the GigaChat assistant for personal and work tasks, and the model weights were open-sourced on HuggingFace and GitVerse under the MIT License — meaning developers anywhere, in principle, can download and build on it directly, not just access it through Sber’s own interface.

What Independent Testing Actually Found

Sber’s own claims are one thing. What happens when someone outside the company actually tests the model against others is a different, more useful data point — and this is where the picture gets more complicated than the press release suggests.

That gap between marketing claims and independent test results isn’t unique to GigaChat — it’s a pattern across the AI industry broadly. But it’s a useful reminder that “outperforms X on our internal tests” and “outperforms X in independent third-party testing” are two very different claims, and only one of them has actually been verified here.

Why a Bank Is Building This

It’s a fair question why Russia’s largest bank — not a dedicated AI lab — is the company behind the country’s leading generative AI model. The answer traces back to Sberbank’s own transformation over the past decade: under CEO German Gref, the bank has deliberately repositioned itself as a broad technology company that happens to also do banking, investing heavily in AI, cloud infrastructure, and consumer devices well beyond its traditional financial services core.

GigaChat sits inside that strategy, but it also sits inside a larger, explicitly stated national one. Sberbank has framed its AI investment as directly aligned with reducing Russia’s dependence on foreign technology amid Western sanctions — language that mirrors, almost exactly, what Japan’s sovereign AI policy and other national AI strategies around the world use, even though the specific geopolitical circumstances driving each country’s approach differ substantially.

Where GigaChat Sits Against Its Domestic Rival

GigaChat isn’t Russia’s only major AI model. YandexGPT, built by Yandex — often described as Russia’s equivalent of Google — is the other significant domestic contender, and by 2026 the two effectively define the Russian AI chatbot market between them. Sber has also released a distinct assistant called Alice, and one independent review specifically flagged that GigaChat has, at points, actually trailed Alice on certain cost-performance comparisons — an unusual dynamic where two products from related ecosystems compete against each other as much as against foreign alternatives.

Both GigaChat and YandexGPT share a structural advantage foreign models don’t have inside Russia: ChatGPT is not officially available in Russia, and much of Western AI tooling is either blocked or operates in a legal grey area requiring workarounds. That’s not a reflection of GigaChat’s technical quality — it’s a market condition that guarantees a captive domestic audience regardless of how the model actually performs against international competitors.

How It Actually Compares Internationally

Here’s where the picture requires some care. Sber’s own comparison points to DeepSeek 3.2 — the Chinese open-weight model — as the benchmark GigaChat 3.5 Ultra “comes close to” on several metrics, while being significantly smaller. That’s a specific, checkable claim, and it places GigaChat’s stated ambition clearly: not competing with the very largest frontier labs’ flagship models, but positioning itself as a credible, efficient open-weight alternative in the same tier as strong open-source Chinese models.

Whether that positioning holds up depends heavily on which numbers you trust — Sber’s own benchmarks, which show competitive results, or the independent testing that placed it in the bottom half of a much broader 47-model comparison. Both can be true simultaneously: a model can perform well on the specific tasks and languages it was optimised for, while performing less well across the more general workload categories that a third-party evaluation is more likely to include.

My Take — Mr Wangdoo

The detail that stands out most to me isn’t GigaChat’s parameter count or its benchmark scores — it’s the decision to open-source the model weights under the MIT License. That’s a meaningful strategic choice, not a technical footnote. A closed, hosted-only model competes purely on how good it is. An open-weight model competes on adoption — how many developers, researchers, and companies build things on top of it, inside Russia and potentially beyond it. Sber picking the open route for its largest model to date suggests it’s playing a longer, ecosystem-building game rather than trying to win a head-to-head capability contest against OpenAI or Anthropic.

I’d also flag the gap between Sber’s performance claims and the independent test results as the single most useful thing in this whole story for anyone actually evaluating the model rather than just reading about it. A vendor’s own benchmarks are marketing material dressed up as data — worth noting, never worth trusting on their own. The mysummit.school review, whatever its limitations, is the kind of arm’s-length testing that tells you something closer to how a model actually behaves on unfamiliar tasks, and its verdict — strong in strategic analysis, weak in team management, unreliable for financial or legal specifics — is a far more useful picture than “GigaChat 3.5 Ultra outperforms our previous flagship.”

The broader story here isn’t really about whether GigaChat is a good model in some abstract sense. It’s that Russia, operating under sanctions and cut off from a meaningful slice of the global AI supply chain, has still managed to ship a 432-billion-parameter model and open-source it within three years of ChatGPT’s original release. Whether that model competes with the frontier is a separate question from whether the underlying capability to build it at all exists — and the answer to that second question is clearly yes.

Frequently Asked Questions

Clayton Samuel (Mr Wangdoo), QFA

Founder and editor, Wangdoo.com. Qualified Financial Adviser with a background in electronics, web development, and cloud infrastructure. This article is based on company announcements and independent technical reporting. Wangdoo has not independently tested GigaChat; no product is promoted.