The Nebius Research Grants Program supports cutting-edge research in AI and AI-driven science by providing compute resources to researchers. Our goal is to boost open science, and open source, and help develop ambitious projects that would otherwise be limited by infrastructure. While we welcome proposals from across the AI ecosystem, we focus our support on areas that align with our research priorities and expertise. Before applying, we encourage you to review the priorities below to determine whether your project is a good fit for the program.

What Makes a Strong Proposal

We generally prioritise projects that:

• Contribute openly to the research community through open-source software, datasets, models, or benchmarks.

• Address significant scientific or technological problems with the potential for broad impact.

• Align with Nebius' research priorities in AI systems, agentic AI, physical AI, life sciences, mathematics, and related fields.

• Offer opportunities for collaboration with Nebius researchers, including technical exchange, or co-authored publications where appropriate.

• Make effective use of compute resources, with a clear plan for how access to our infrastructure will accelerate the research.

• Intend to publish results in peer-reviewed venues or otherwise contribute findings back to the research community.

• Engage with the research community through Nebius Science and other Nebius media channels. These may include interviews, case studies, feature articles, webinars, etc.

Areas We Actively Support

Inference Optimization

Research on making large-model serving faster, cheaper, and more scalable on modern GPU clusters. We welcome proposals on speculative decoding, KV-cache compression and management, low-precision inference (FP8/FP4 quantization), sparse and long-context attention, disaggregated prefill/decode architectures, MoE serving, and kernel- or scheduler-level optimizations. Strong projects pair algorithmic ideas with real systems benchmarks — latency, throughput, and cost per token at scale.

Post-Training

Research on turning pretrained models into capable, aligned, and reliable systems. Relevant directions include reinforcement learning from human or AI feedback, RL with verifiable rewards, reward modeling, preference optimization methods (DPO and beyond), reasoning and chain-of-thought training, synthetic data generation, distillation, and rigorous evaluation of post-trained models. We especially value work that improves sample efficiency or makes post-training pipelines more reproducible and open.

We welcome research on building the environments, tasks and rewards that RL post-training and agent training depend on: executable, verifiable, and realistic. Please apply with proposals on environment and task synthesis, sandboxed execution infrastructure for high-throughput parallel rollouts, verifiable and outcome-based reward design, robust verifiers and reward-hacking detection, contamination- and leakage-resistant task construction, long-horizon and multi-step environments (software engineering, web and computer use, research automation).

Agentic Systems

Research on LLM-based agents that plan, use tools, and act over long horizons. Topics include multi-agent orchestration, tool-use and code-execution agents, memory and long-horizon planning, RL environments and training for agents, agent benchmarks and evaluation, and reliability/safety of autonomous systems. Proposals should target measurable improvements on realistic tasks (e.g., software engineering, research automation) rather than demos alone.

Multimodal, Omni, World, Physical AI, Science, and Healthcare Models

Research on models that perceive, reason about, generate, and act across multiple modalities and real or simulated environments. Topics include unified omni-models spanning language, vision, audio, video, and sensor data; world models for prediction, simulation, and planning; vision-language-action (VLA) models, embodied models for robotics and physical interaction; large-scale simulation and synthetic data generation, sim-to-real transfer, reinforcement learning for robot policies; domain-specific models for science, healthcare, engineering, and other specialized applications; and associated tools for data generation, training, evaluation, deployment, and safety.

Strong proposals should demonstrate measurable gains on realistic, multimodal or embodied tasks and clearly address data efficiency, generalization, latency, robustness, and real-world evaluation. They treat data and simulation pipelines as core research contributions, not just infrastructure and validate on real hardware or established benchmarks instead of curated demos. Typically training from scratch is out of scope unless particularly efficient, but techniques spanning multiple architectures or applications are of interest.

AI Portability

Research on making models, agents, skills, memory, and workflows transferable across providers, runtimes, and agent harnesses without sacrificing capability or creating dependence on a single proprietary stack. Topics include model routing and replacement, cross-model and cross-harness transfer, portable agent profiles and action interfaces, standardized trajectory and memory formats, interoperable tools and protocols, and evaluation of model or harness changes. We especially value proposals that enable migration from closed to open models, measure quality, cost, and latency tradeoffs on realistic tasks, and produce reproducible benchmarks, open standards, or reference implementations.

Novel Architectures and Model Design

Research on model architectures that could improve capability, efficiency, or long-context behavior beyond standard Transformer scaling. Topics include attention alternatives, sparse/linear/latent attention, state-space and recurrent models, memory-augmented architectures, dense-to-MoE conversion or distillation, multi-token or diffusion-style prediction, thought-token / latent-reasoning mechanisms, and architectures optimized for tool use or long agent traces. Strong proposals should include realistic baselines and a path to evaluating quality-cost tradeoffs, not only synthetic benchmarks.

If you’re working on research that fits our mission and would like to join the team, explore our open roles