How they use Nebius and Tavily

Nebius named JetBrains as a managed client in 2024. In June 2026, Nebius said it supplied part of the training infrastructure for Mellum2, JetBrains' coding model. In April 2026, JetBrains described working with cloud providers including Nebius for latency-sensitive IDE code completion with regional presence in the US, Europe and Asia Pacific. JetBrains Academy also partnered with Nebius on free AI-assisted programming courses using Junie.

JetBrains uses Tavily in its developer tools. Its provider documentation names Tavily. AI Assistant integrates Tavily's API directly into JetBrains IDEs for current web answers and documentation searches scoped to framework sites such as PyTorch and NumPy. Tavily says its built-in safe search, formatting and domain filters saved months of engineering effort. It reports that complaints about outdated or missing answers disappeared after integration. Alexandra Charikova said the collaboration went live in spring 2025. Tavily added crawling support at JetBrains' request. Its AI Assistant account says JetBrains is considering further opportunities for Tavily in the platform.

Junie integrates Tavily as a native web-search tool for current API and niche-library documentation with citations. JetBrains chose model-agnostic retrieval of relevant content chunks. Tavily's reranking limits context use. Tavily says integration required little engineering overhead and moved quickly into validation. JetBrains defined search triggers for questions beyond the model's knowledge cutoff, niche libraries and open-ended library comparisons using current documentation, recent benchmarks and community discussion. Nick Frolov describes routing between the codebase, agent memory and external search. He says Tavily's HTML stripping and controllable chunking help reduce token costs. A Tavily presenter reports searches taking roughly 1.5–2 seconds in Junie's implementation. Tavily reports reduced hallucinations about recent APIs. JetBrains Head of Product Nick Frolov described real-time web search correcting hallucinated library methods while building a Telegram integration. Tavily works in Junie without configuration, including through the Junie CLI. Its customer story reports 10–20 minutes saved per avoided wrong-API debugging cycle. These Tavily integrations are separate from the disclosed Nebius model-training infrastructure.