Russia's Sber Releases Open-Source MoE LLM GigaChat 3.5 Ultra

Sber, Russia's largest bank and a major AI developer, has released GigaChat 3.5 Ultra, the latest iteration of its proprietary large language model series. The new version adopts the Mixture of Experts (MoE) architecture, following a design trend increasingly common among state-of-the-art models.

GigaChat 3.5 Ultra represents a significant upgrade within the GigaChat family, with the "Ultra" designation suggesting a substantial increase in parameter count and capability. The MoE architecture enables the model to activate only a subset of expert networks per inference, maintaining high capacity while optimizing computational cost.

The open-source release strategy allows global developers and researchers to freely download, deploy, and fine-tune the model. This aligns with the broader industry movement toward open-weight models, though it also raises questions about how Sber's approach compares to other open-source LLMs from Chinese and Western developers.

The release is notable given the geopolitical context. Despite tightening export controls on advanced AI chips to Russia, Sber has continued to iterate on its model series, suggesting a degree of domestic AI supply chain resilience. GigaChat has previously demonstrated strong performance on Russian-language NLP benchmarks.

MoE has become a dominant architectural choice in the LLM space through 2025-2026, with models like DeepSeek-V3, Qwen2.5-MoE, and Mixtral all adopting variants of the approach. GigaChat 3.5 Ultra adds another data point to this architectural convergence.

Sber has not yet released a full technical report or detailed benchmark results for the new model. The developer community will be watching for performance data, particularly on multilingual tasks and reasoning benchmarks, to gauge how GigaChat 3.5 Ultra stacks up against comparable open-source alternatives. Attention is also focused on the model's licensing terms and whether it permits commercial use.