Sber open-sources GigaChat 3.5 Ultra, a 432-billion-parameter model

Sberbank has open-sourced its new flagship model, GigaChat 3.5 Ultra. It has 432 billion parameters, almost 40% fewer than the previous flagship, GigaChat 3.1 Ultra, which had 700 billion. The model weights are available for developers to deploy themselves, without going through the company's closed API.

How much smaller is the model?

GigaChat 3.5 Ultra has 432 billion parameters, compared with 700 billion for GigaChat 3.1 Ultra — a reduction of almost 40%. The developers stress that this is not a "smaller but cheaper" trade-off: they say new data, an updated training recipe and architectural changes made the model both more compact and stronger.

• GigaChat 3.5 Ultra — 432 billion parameters, released as open source
• GigaChat 3.1 Ultra (the line's previous flagship) — 700 billion parameters
• Reduction in model size — roughly 40%, or 268 billion parameters
• For the first time in the GigaChat line, the in-house hybrid architecture has been scaled to hundreds of billions of parameters

What changed in the architecture

For the first time, the Sber team scaled GigaChat's in-house hybrid architecture to hundreds of billions of parameters in a single release. In models like this, a hybrid architecture usually means a mix of different layer types and attention mechanisms, which lets developers scale up without a proportional rise in compute costs. The developers say this allowed them to speed up inference (response generation) and expand the model's capabilities at the same time — primarily in code, agentic scenarios (where the model calls tools and carries out chains of actions on its own) and complex domains.

"This is not a 'smaller but cheaper' compromise: thanks to new data,

an updated training recipe and architectural changes, the model has become stronger, as well as more efficient in memory use and generation speed," Sber wrote in its official blog on Habr.

Why did the company shrink the model?

Sber frames the smaller size as an engineering decision rather than a cost-cutting move: a smaller model uses memory more efficiently and generates faster while maintaining or improving quality. The 268-billion-parameter gap between GigaChat 3.1 Ultra and GigaChat 3.5 Ultra has a direct impact on infrastructure: the fewer parameters a model has, the less video memory it needs to load and the faster it generates tokens on the same hardware.

What does this mean for developers?

For teams that deploy open source models themselves, shrinking the flagship from 700 billion to 432 billion parameters lowers the hardware barrier to entry while delivering comparable or better response quality. GigaChat 3.5 Ultra targets code and agentic scenarios, among other things — exactly the tasks where developers most often compare open source models against competitors' closed APIs.

What it means

Sber continues to release its flagship GigaChat models openly while shifting toward more compact and faster architectures. That mirrors a broader trend across the large language model industry, where developers are trying to balance model size, infrastructure costs and practical usefulness in code and agentic scenarios.