Yandex built Alisa AI's research agent without fine-tuning weights and cut GPU usage severalfold
· Source: original
⚡ Alisa AI's Deep Research agent was built without fine-tuning the weights
Yandex rolled out a research agent as a separate "Research" mode in the chat with Alisa AI. The model weights weren't touched at all — the team managed without fine-tuning, and GPU 🚀 consumption dropped severalfold. The team's breakdown is on habr.com
What the agent does in action: builds a unique plan for solving a task, runs hundreds of searches per query, visits websites — including pages with dynamic JavaScript content, writes and executes Python code for complex calculations, and works with downloaded files.
An example query from the article: "Plan me a two-week road trip to Dagestan by car with kids."
The agent went from first prototype to production in a year. Along the way: dozens of layers of scaffolding and one discarded prototype.
The team lead for the "Research" mode is Prokhor. The product side was led by the agent's product manager, Ruslan Iliev: he formulated the product goals, defined the toolset and the validation query set, and ran the launch from a closed waitlist to 100% production.
The material is dated September 14, 2026; the breakdown is in the digest
🤖 Interested in AI agents and automation?
Prompts for building AI agents and automations — read up on the topic:
🔗 The entire prompt library · "AI Agents" category
A ready-made product on the topic: AI Agent Skills Pack — grab it and apply it right away.