Today I went to Sber’s AI R&D Day conference for research teams and people building AI systems. Here are the things that stood out.

I enjoyed the talk about compact language models. Large LLMs can reproduce substantial passages from the Harry Potter books, which is certainly impressive. But for many practical tasks, that breadth of knowledge is more than you need. Answering accurately from the context provided, and recognizing when there isn’t enough information, is far more useful.

The AIRI team presented Optimal Cognitive Core, a family of small models being developed with exactly this focus. According to their tests, OCC-RAG, with 1.7 billion parameters, outperforms much larger models on a number of tasks involving reasoning over context and keeping answers faithful to sources. I’ve used similar small language models from Microsoft on my modest MacBook Air M1, and was pleasantly surprised by the results.

Access to GigaCode Universal Agent opened alongside the conference. Its command-line interface looks very much like Claude Code. I’ve installed it and am still finding my way around.

I also got to talk to the head of development about agent security. I asked which approach Sber was taking: bundled, with security provided by the agent platform vendor, or unbundled, with protection supplied separately by other developers. He framed their choice differently: there is a spectrum between complete protection and none at all, and the team has chosen a particular point on it. Within GigaCode, they build security in several layers: instructions for the model, constraints in the harness, and a separate judge model that checks actions. A sensible approach in principle.

The recently released GigaChat 3.5 Reasoning also came up repeatedly. Sber presents it as the first Russian model with a full reasoning mode and open weights. The earlier Sber models hadn’t quite met my research needs, so I’m waiting for a chance to try the new one. It isn’t available through Sber’s API yet, but Cloud.ru has promised access on September 21.

Overall, I came away with a positive impression. I particularly liked the opening talk on AI trends. Once the recording is available, I’d recommend watching it.