The role

Nebius Token Factory is looking for scientists who can turn frontier LLM and VLM inference bottlenecks into research questions, conduct rigorous experiments, and translate their findings into production capabilities. You will own focused research and optimization projects, write strong code, and produce prototypes, publications, and technical artifacts that engineers can build on.

Your research will span quantization, distillation, speculative decoding, KV-cache optimization, and model/runtime co-optimization, alongside inference engines and distributed inference architectures. You will investigate how these approaches interact, evaluate their trade-offs across model quality, latency, throughput, memory footprint, and cost per token, and work with engineering teams to bring promising results into production.

Your responsibilities

• Lead research projects in efficient LLM and VLM inference, from hypotheses and experiments through ablations, prototypes, and production handoff.
• Develop and evaluate methods spanning quantization, quantization-aware training (QAT), distillation, speculative decoding, KV-cache reuse, KV-cache compression, long-context inference, MoE routing, and model/runtime co-optimization.
• Build prototypes using PyTorch, Triton, CUDA-adjacent tooling, or inference-serving frameworks, and collaborate with MLEs and platform engineers to turn them into production components.
• Research LLM request routing and scheduling strategies, including cache-aware load balancing and prefill–decode disaggregation (PDD). Evaluate how queueing, KV-cache transfer, worker placement, and prefill/decode capacity allocation affect latency, throughput, and serving cost.
• Investigate multi-node inference for dense and mixture-of-experts models, including wide expert parallelism (WideEP). Study expert placement, load imbalance, and computation/communication trade-offs, and use the findings to guide model and system architecture choices.
• Develop rigorous evaluations covering quality, latency, throughput, numerical stability, memory footprint, tail latency, and cost per token. Compare serving architectures under consistent workload conditions and GPU budgets.
• Work with MLE, GPU kernel, backend infrastructure, product, and customer teams to select research priorities with measurable production impact.
• Share results through internal reports, technical blogs, papers, and open-source artifacts, and mentor engineers and scientists on experimental design, scientific rigor, and model/system trade-offs.

Must-haves

• A PhD in computer science, machine learning, ML systems, computer systems, computer architecture, electrical engineering, applied mathematics, or a closely related discipline.
• A strong publication record or equivalent research artifacts in ML, ML systems, efficient inference, model compression, quantization, distillation, serving systems, or related areas.
• Strong Python and PyTorch implementation skills, with the ability to turn ideas into experiments and working prototypes.
• Deep knowledge of LLMs, VLMs, transformer inference, decoding algorithms, model compression, quantization, and production-serving trade-offs.
• Strong experimental design skills covering ablations, baselines, metrics, statistical reasoning, and failure analysis.
• Excellent written and verbal communication.

Nice-to-haves

• First-author publications at venues such as NeurIPS, ICML, ICLR, MLSys, ACL, EMNLP, ASPLOS, OSDI, SOSP, ISCA, or HPCA.
• Experience deploying ML models or inference optimizations in production.
• Experience with vLLM, SGLang, TensorRT-LLM, NVIDIA Dynamo, FlashAttention, FlashInfer, Triton, CUDA, or PyTorch internals.
• Experience applying post-training, SFT, DPO, RLHF, RLAIF, preference optimization, or synthetic data generation to inference quality or efficiency.
• Open-source research artifacts, widely used benchmarks, technical blogs, or invited talks demonstrating contributions to efficient AI systems.
• Research or implementation experience in one or more areas of distributed inference system architecture, such as LLM request routing, PDD, multi-node serving, or MoE expert parallelism, including WideEP.