Empirical telemetry benchmarking Russian foundation models (YandexGPT 4, GigaChat MAX), tech conglomerates (Yandex, Sber, AIRI), and supercomputers.
| # Rank | Model Name | Developer | Architecture / Params | MMLU-RU Score | License |
|---|---|---|---|---|---|
| #1 | YandexGPT 4 (Pro/Lite) ↗ | Yandex | Dense / MoE (~120B+) | 96.4 / 100 | Proprietary Cloud API |
| #2 | GigaChat MAX ↗ | Sber AI | MoE (~140B+) | 95.8 / 100 | Proprietary Cloud API |
| #3 | Kandinsky 3.1 ↗ | Sber AI / AIRI | 3.3B Diffusion | 94.5 / 100 | Open-Weights Commercial |
| #4 | Yandex ART ↗ | Yandex | Diffusion Transformer | 94 / 100 | Proprietary Cloud API |
| #5 | YaFSDP ↗ | Yandex | Open Source Tool | 98.2 / 100 | Open Source (Apache 2.0) |
| #6 | DeepPavlov Dream ↗ | MIPT / DeepPavlov | Open-Weights NLP | 91.2 / 100 | Open Source (Apache 2.0) |
| #7 | T-Bank T-Lite 8B ↗ | T-Bank AI Research | 8B Dense | 92.8 / 100 | Open-Weights Commercial |
| # Rank | Enterprise / Lab | City | Primary Domain | Valuation / Scale | Moat Rating | RSS Score |
|---|---|---|---|---|---|---|
| #1 | Yandex (Яндекс) ↗ | Moscow | Foundational LLMs & Cloud | $12,500M | Deep Sovereign Moat | 97.8 |
| #2 | Sber (Sber AI) (Сбер (Sber AI)) ↗ | Moscow | Fintech AI & Robotics | $38,000M | Deep Sovereign Moat | 96.5 |
| #3 | AIRI (AI Research Institute) (Институт AIRI) ↗ | Moscow | Frontier AI Research | $120M | High Moat | 95.2 |
| #4 | Skoltech (CDISE) (Сколтех) ↗ | Moscow (Skolkovo) | Frontier AI Research | $250M | High Moat | 93.8 |
| #5 | T-Bank AI Research (Tinkoff) (Т-Банк ИИ) ↗ | Moscow | Fintech AI & Fraud Modeling | $7,500M | High Moat | 94 |
| #6 | Cognitive Pilot (Когнитив Пилот) ↗ | Moscow / Tomsk | Autonomous Navigation | $180M | Deep Sovereign Moat | 92.4 |
The fastest supercomputer in Russia and Eastern Europe, driving Yandex search neural ranking and YandexGPT distributed pre-training.
High-throughput Yandex supercomputing node dedicated to multimodal neural search, Yandex ART video rendering, and voice processing.
Sber's flagship sovereign supercomputer powering GigaChat training, enterprise ML sandboxes, and financial risk simulations.
Dedicated commercial AI supercomputer for telecom voice analytics, enterprise computer vision, and retail LLM agents.
| Role Title | Seniority Tier | City Hub | Monthly RUB | Annual RUB | Annual USD Equiv |
|---|---|---|---|---|---|
| Lead / Principal LLM Pre-Training Scientist | Staff / Principal (10+ Yrs) | Moscow | ₽650,000 | ₽7.8M / yr | $86,000 USD |
| Senior AI Alignment & Post-Training Engineer (RLHF) | Senior (6-9 Yrs) | Moscow | ₽450,000 | ₽5.4M / yr | $60,000 USD |
| Senior Computer Vision & Diffusion Architect | Senior (6-9 Yrs) | Saint Petersburg | ₽400,000 | ₽4.8M / yr | $53,000 USD |
| Middle+ ML / High-Load Backend Engineer | Mid (3-5 Yrs) | Russia Remote | ₽280,000 | ₽3.36M / yr | $37,000 USD |
| AI Edge & Embedded Robotics Engineer | Mid (3-5 Yrs) | Novosibirsk | ₽260,000 | ₽3.12M / yr | $34,500 USD |
The premier Russian foundation models are YandexGPT 4 (by Yandex, scoring 96.4/100 on MMLU-RU), GigaChat MAX (by Sber AI, scoring 95.8/100), Kandinsky 3.1 (diffusion vision network by Sber/AIRI), and YaFSDP (open-source distributed LLM training framework accelerating pre-training by up to 26%).
The fastest sovereign supercomputer in Russia and Eastern Europe is Yandex's Chervonenkis (29.41 PFLOPS Rpeak, 1,592 NVIDIA A100 nodes), followed by Yandex's Lyapunov (16.74 PFLOPS Rpeak), Sber's Christofari Neo (14.90 PFLOPS Rpeak), and MTS's MTS GROM (2.26 PFLOPS Rpeak).
Lead LLM pre-training scientists in Moscow earn a median salary of ₽650,000/month (~$86,000 USD/year), while Senior RLHF & alignment engineers earn ₽450,000/month (~$60,000 USD/year), and Middle+ ML engineers earn ₽280,000/month (~$37,000 USD/year).
AKI benchmarks Russia's sovereign AI readiness at 88.4 / 100 (Tier 1 Sovereign Resilience), reflecting high algorithmic and search independence, strong domestic foundation models, and an active aggregate supercomputing capacity exceeding 63.3 PFLOPS.