Parallel Frontier
CHINA & RUSSIA AI ECOSYSTEMS, READ IN THE ORIGINAL
ISSUE Nº 2 · RUSSIA BEAT
Russia Beat · Lab-Watch-Led

From Scratch, For Whom?

GigaChat 3.1 Ultra and T-pro 2.0 are real domestic builds. The story this quarter isn't the capability gain. It is the distribution strategy.

Three Russian releases anchor this quarter, and the first thing to say about them is that they are real. Sberbank's GigaChat 3.1 Ultra is an open mixture-of-experts model, roughly 702 billion total parameters with about 36 billion active, trained from scratch on Russian data and posted to Hugging Face in November 2025; a reasoning mode followed in March 2026, which a management-scenario review found worsened results where precise calculation was needed. T-Bank's T-pro 2.0 is the more interesting research contribution: open-weight, built around a Cyrillic-dense tokenizer that lifts the share of Russian words encoded in two tokens or fewer from about 38 percent to about 60 percent, and shipped with an instruction corpus, a math benchmark, and speculative-decoding weights that make it reproducible, which most Russian releases are not. Yandex is the strongest technical player of the three, with its Alice AI family, YandexGPT 5.1 Pro, and an open 8-billion-parameter pretrain, but it is not the state's chosen flag-carrier, a function of its complicated relationship with the Kremlin.

This issue opens with those specifics on purpose. Trained from scratch, no foreign base weights, genuinely open: that much survives contact with the primary sources, and the credibility of everything after it depends on not treating Russian AI as a punchline.

What the sovereignty narrative around these builds does not say is who the pitch is for, and what it quietly depends on. That gap, not the benchmarks, is the subject of this issue.


Key Judgments
1
High confidence
We judge it almost certain that GigaChat 3.1 Ultra and T-pro 2.0 are real domestic builds, not relabelled foreign weights. GigaChat is an open mixture-of-experts model (~702B total, ~36B active) trained from scratch on Russian data; Sber released it and its sibling models with full weights, announced by Herman Gref at AI Journey 2025. T-pro is open-weight, with a Cyrillic-dense tokenizer. For these two, "from scratch" is substantively true. (Confidence: high.)
2
High confidence
We judge it likely that the significance this quarter is not the capability gain. It is the distribution strategy. Sber is pitching sovereign AI to Latin America, Africa, and Asia, converting a capability deficit into a values proposition, and Putin's keynote gave it a state frame: the Russia-led International AI Alliance grew to 28 organisations in November 2025, adding Brazil, Chile, Egypt, Kenya, South Africa and others. (Confidence: moderate to high.)
3
High confidence
We judge it almost certain that the binding constraint is still compute and price-to-quality, not architecture. On cost-adjusted quality, the flagship Russian models sit below the leading Chinese open models by multiples, and Putin's keynote concedes the constraint at state level: compute and energy are named the mainline of investment, with 38 planned reactors and small nuclear sited at data centres. (Confidence: high.)
4
High confidence
We judge it likely that the July 2026 softening of Russia's AI law followed a deregulatory line Putin set himself. The State Duma passed the law on 8 July 2026, creating "sovereign" and "national" model categories; the revision that cut it from 21 articles to 13 dropped the Russian-only-data and Russian-citizens-only requirements, letting "national" models use foreign open-source components. Putin's keynote had already called to strip barriers to sovereign tech, and Russian coverage ties the change to the fact that much Russian AI runs on open-source foreign bases like DeepSeek. The constraint, not political will, set the ceiling. (Confidence: high.)
5
Low confidence
We judge it roughly even whether any Global South state announces a GigaChat-based national deployment within two quarters of the June pitch (by roughly early December 2026). We hold this out less as a forecast than as a pre-committed indicator: if it fires, the export ambition has become a transaction; if the window closes empty, the likelier read is that values-alignment rhetoric is outrunning what sanctioned compute and thin non-Russian data can deliver. (Confidence: low.)

Lead Analysis
From scratch, for whom? Sovereignty repackaged as an export pitch

What the record is. The quarter's Russian flagship releases, read from primary sources: GigaChat 3.1 Ultra, an open mixture-of-experts model (~702B total parameters, ~36B active) trained from scratch on Russian data and released on Hugging Face in November 2025; and T-pro 2.0, open-weight, with a Cyrillic-dense tokenizer. Behind them stands Putin's keynote at AI Journey 2025 (Sber headquarters, 19 November 2025), which framed the whole program as a matter of state, technological and, in his phrase, values sovereignty. The verbatim passage: «Критическую зависимость от чужих систем мы не можем допустить. Для России это вопрос государственного, технологического и, я бы сказал, ценностного суверенитета.» ("We cannot permit critical dependence on others' systems. For Russia this is a question of state, technological and, I would say, values sovereignty.") In the same speech he said that in certain areas, partly public administration and partly the work of the special services and force agencies, Russia must rely exclusively on its own base developments.

The following separates the record from my reading.

The fact register. The releases are real, and saying so is the point: genuine domestic builds, not relabelled foreign weights. At AI Journey 2025 Herman Gref announced the open release, with full weights, of the flagship (GigaChat 3.1 Ultra, then in preview) and its siblings, GigaChat Lightning, the GigaAMv-3 speech model, and the Kandinsky and Kandinsky 5.0 image and video models, calling it the largest open-source project in Europe. And the benchmarks undercut the capability claim: on cost-adjusted quality the flagship Russian models lag the leading Chinese open models, and this is sourced, not asserted: Sber's own Anton Frolov, who heads its generative-AI development, treats DeepSeek openly as proof that frontier capability no longer needs a trillion-dollar budget.

The political read. The narrative around the releases is sovereignty, and the labs' "no foreign base weights" framing is the technical answer to a political instruction. When the capability claim fails, the pitch shifts ground rather than folding: in a Reuters interview ahead of the St Petersburg forum in June 2026, Sberbank First Deputy CEO Alexander Vedyakhin framed the product as slower and not as smart as Anthropic, Grok, or DeepSeek, but aligned with your values, sold to states that want sovereign AI and cannot afford to build it.

The economic read. Read together, the move is rational. A losing capability position is repositioned as a values-and-residency offer, mirroring China's own Global South open-source play; Putin's keynote supplies the state-level version, an International AI Alliance now 28 organisations strong and a proposal to harmonise AI law with partners through BRICS, the SCO and the EAEU. "Sovereignty" stops being a domestic shield and becomes an export pitch. The irony underneath it is the dependency the pitch omits: Chinese chips, and DeepSeek as the template everyone copies.

What it means / what's next. For Russia-watchers: judge the program by distribution deals, not benchmarks: the product is sovereignty, and its market is states. For Global South ministries: the pitch's honest content is residency and control, not capability; the dependency it omits prices the sovereignty on offer. For the China–Russia desk: Sber's chip-sourcing is the standing item, and the narrative's silence about it is the tell. Indicators to watch: (1) any Global South state announcing a GigaChat-based national deployment within two quarters, the tripwire in Key Judgment Five; (2) whether the Russian-only-data requirement dropped from the sovereign-AI bill before its July 2026 passage returns in secondary regulation (Key Judgment Four); (3) Sber's chip-sourcing for GigaChat; (4) whether T-pro 2.0's corpus, benchmark and tokenizer work is picked up outside T-Tech: reproduction is the compliment that counts.


内外有别 The Translation Gap
The pitch at home and the pitch abroad

The standing section, on the Russia beat this time: the same discipline Issue 1 applied to Chinese state text, applied to the two registers of Russian sovereign-AI talk, the one aimed inward, and the export English aimed at everyone else.

The fact. The program speaks two scripts. Inward: Putin's AI Journey 2025 keynote (critical dependence on foreign systems "unacceptable"; a matter of state, technological and values sovereignty; own base developments only in administration and the security agencies) and the paired vocabulary written into the 8 July 2026 AI law, суверенный ИИ ("sovereign AI," fully Russian, the true "from scratch") and национальный ИИ ("national AI," Russian-controlled but permitted foreign open-source components). Outward: the June 2026 export pitch: slower, not as smart, but aligned with your values, resident on your soil.

What crosses over. The offer: sovereignty as a product. Values alignment, data residency, independence from the American stack, a capability deficit converted into a values proposition for states that cannot build their own.

What stays in Russian. The register that reveals what the product is for:

The read. We assess the export pitch is the domestic mandate translated into marketing. The deficit conceded abroad ("slower, not as smart") is unsayable at home, where the keynote frames sovereignty as security; and the dependency underneath is unsayable in both. Each register omits what the other requires.

The honest caveat. The two quotations doing the load-bearing work are confirmed: Putin's sovereignty passage against the Kremlin transcript of the AI Journey 2025 keynote, and Vedyakhin's concession against the Reuters interview. Its deregulatory direction is confirmed against the keynote, and the bill's passage (State Duma, 8 July 2026) and the 21-to-13 cut check out across the Russian coverage. What remains genuinely forward-looking is the export tripwire in Key Judgment Five.

Why it matters. The ministries this pitch targets will read it in English. What the Russian register adds is the pitch's function: compliance with a political instruction, repackaged for export. Reading both is the difference between evaluating a product and evaluating a promise.


The Docket, instruments behind Russia's sovereign-AI pitch

Lab Watch, GigaChat 3.1 Ultra, T-pro 2.0, and the Yandex family against DeepSeek, the template the field measures itself by

Terminology Watch, суверенный ИИ beside национальный ИИ

Signals
Read in the original. Verify before you cite. Parallel Frontier is written by Caroline Swartz. Chinese language & politics (NYU), Russian politics MSc (KCL), Mandarin SIGINT spanning defensive and offensive cyber operations, security operations leadership, short- and long-form reporting in five languages, AI safety research, and an MA in computational linguistics in progress (UCL).