Sber has introduced GigaChat 3.5 Ultra – a new flagship model for code, mathematics, long texts, and AI agents. It is already available for free in the GigaChat assistant and open access for developers.
The main difference is linear attention technology. A regular neural network re-checks the entire previous text with each new word. GigaChat 3.5 Ultra gradually collects context and keeps it in memory – therefore, according to Sber, it processes long documents up to four times faster.
The model is expected to work more accurately with contracts, reports, and technical regulations, as well as write and check code, perform calculations, and execute multi-step tasks. It can be used as a basis for AI agents that independently search for data, access services, and return a ready result.
The new version is almost twice as compact as the previous GigaChat Ultra and requires fewer resources. This should allow companies to run it on more affordable hardware, rather than just expensive computing clusters.
Read more on the topic:
- Text as an Image: Sber Intern Releases AI Faster Than GigaChat
- ChatGPT is being disabled in Russia: Experts advise switching to domestic AI
- Sber launches AI servers for lease: GigaChat Enterprise received for 36 months with no upfront costs