Sberbank released GigaChat 3.5 Ultra, a 432-billion-parameter mixture-of-experts model and the largest in Sber's lineup to date, open-sourcing the model weights on both HuggingFace and the Russian platform GitVerse under the MIT license. It follows GigaChat 2, launched in February 2026, which Sber said scored 80.46 on the MMLU-RU benchmark -- a figure the company reported as ahead of GPT-4o and Claude 2 on Russian-language tasks specifically (a claim from Sber's own benchmark reporting, not independently verified here).
Separately, Yandex has deployed YandexGPT 5, which reportedly matches GPT-4o's performance on 64% of tasks, with a large context window and token-based ruble billing built directly into Yandex's own ecosystem. Sber has also built Kandinsky 5.0, an image and video generation model positioned against international competitors.
The consistent theme across both companies is real software-side progress running on a constrained hardware ceiling: semiconductor export sanctions limit Russian labs' access to the frontier chips that US and Chinese labs train on, so the competition here is happening on model efficiency and quality within that ceiling, not on trying to out-scale it.
This is the first Lab & Nation Watch entry on Russia because there was genuinely sourced material this week (a 432B open-weight release, a benchmark claim against GPT-4o) -- not because Russia gets a standing weekly slot; expect this entry to be absent in weeks without comparable substance.