We build infrastructure that underpins cutting-edge research by Nebius AI R&D.

⚛️ What is Nebius AI R&D?

Nebius AI R&D is a team focused on open, applied AI research. We work at the intersection of research and real-world deployment, building open benchmarks, datasets, and training methods that push the frontier of agentic systems.

Examples of our work include:

• SWE-rebench-V2 — The largest open dataset in the world for training coding agents
• swe-rebench.com — A continuously updated benchmark for evaluating coding agents in realistic software engineering tasks.
• Applying reinforcement learning for agent training in long-context multi-turn scenarios
• Developing novel loss functions to accelerate the training of speculators for inference of LLMs

We collaborate closely with Nebius Token Factory, translating our research breakthroughs into tangible product advancements — including fine-tuning services and inference performance optimizations.

⚙ What we do

Modern AI research is extremely infrastructure-heavy: multi-node GPU training jobs, heavy LLM inference deployments, and thousands of agent sandboxes running in parallel.

Our Research Infrastructure team develops and supports all types of infrastructure needed to empower the researchers in their work. We own:

• A global network of GPU-equipped Kubernetes clusters
• An in-house sandbox with Git-like branching for safe execution of AI-generated code: contree.dev
• Large-scale pipelines to run, evaluate, and train AI agents – including our in-house SWE agent implementation
• And much more

Some of these internal research systems evolve into public-facing products — such as ConTree.

🧰 Our stack

• Base infra: Kubernetes, Terraform, Helm, Nebius cloud, Grafana, Keda, Volcano
• ML: Papyrax, vLLM, JAX, Python, Airflow, W&B
• ConTree: QEMU, Btrfs, Go, Python