December 2025 Summaries

Reinforcement learning (RL) showcases advanced results in software engineering tasks but faces significant infrastructure challenges, requiring complex MLOps workflows with asynchronous setups and fine-grained experimentation. In contrast, behavioral cloning and model distillation, which involve supervised fine-tuning, offer simpler alternatives. Rejection fine-tuning (RFT) is particularly effective by using successful trajectories from multiple solution attempts, circumventing the infrastructure demands of RL. The research contributes a dataset named nebius/SWE-rebench-openhands-trajectories, containing 67,074 agent trajectories from 1,823 Python repositories on GitHub, generated by the Qwen3-Coder-480B-A35B-Instruct model using the OpenHands framework. RFT checkpoints at two scales, 30B and 235B, show promising performance on benchmarks like SWE-bench Verified, indicating the potential of RFT in capturing high-quality behavior. The work also provides comprehensive documentation to ensure reproducibility in evaluations using the OpenHands framework, emphasizing the importance of detailed configuration management to avoid data leakage and infrastructure instability.

Bioscience is leveraging open-source tools and cloud technology to advance the understanding of protein complexes crucial for cellular processes, with the University of California, San Francisco (UCSF) playing a significant role in this endeavor. Supported by Nebius AI Cloud and its educational arm, UCSF researchers, including Dr. Benjamin Webb, are updating the Integrative Modeling Platform (IMP) to run on modern GPU cloud platforms, thereby enhancing access and usability for scientists globally. IMP, an open-source software, integrates data from various experimental sources to model large protein complexes, facilitating discoveries in medicine and drug design. This transition from CPU to GPU accelerates the modeling process, allowing for more extensive studies of complex structures like the nuclear pore complex. The initiative aligns with efforts to update and expand the Protein Data Bank (PDB), a vast archive containing over 240,000 entries that support research across molecular biology and biotechnology. The PDB's integration of new models developed with IMP underpins advancements in AI-driven molecular studies, promising transformative tools for understanding biomolecules in their natural environments.

Neural networks and transformers are foundational to modern AI, with neural networks comprising layers of neurons that process information incrementally, while transformers excel at understanding and generating text sequences by focusing on context. Effective AI models rely on embeddings, weights, and quantization to optimize memory usage and enhance performance without compromising quality significantly. The transformer model's strength lies in its ability to relate every word in a prompt to others, enabling coherent and context-aware responses. During inference, models undergo a workflow involving tokenization, embedding lookup, and decoding, efficiently managed through techniques like the KV cache and CUDA graphs. Attention mechanisms play a crucial role by dynamically focusing on relevant parts of the input, and the choice of attention backend can impact performance. Inference efficiency is further enhanced by parallelism, as seen in vLLM, a library that offers high-performance LLM serving through advanced memory management and parallel processing strategies. Model artifacts, including weights and tokenizer assets, are crucial for maintaining consistency across environments, and licenses must be adhered to for legal compliance in AI deployment.

In a significant development for AI infrastructure in Europe, Nebius has launched the first operational deployment of the NVIDIA GB300 NVL72 platform at its expanded Finland data center, marking the region's introduction to NVIDIA's advanced AI technology powered by Blackwell Ultra GPUs and Quantum-X800 InfiniBand networking. This system, consisting of 72 GPUs functioning as a unified supercomputer node, offers remarkable performance enhancements for generative AI, large-scale model training, and advanced inference workloads, delivering up to a 50x boost in AI Factory output and a 30x improvement in real-time video generation compared to previous systems. This advancement allows users to train and optimize complex AI models efficiently without facing bottlenecks, leveraging Nebius AI Cloud's scalability and extensive AI/ML tools. The new infrastructure was unveiled during a data center tour led by Nebius' CEO Arkady Volozh, emphasizing its role in supporting Europe's AI development and innovation.

In the realm of healthcare and life sciences, rapidly establishing a comprehensive AI stack is essential, and this can be efficiently achieved with NVIDIA NIM, a suite of high-performance microservices designed for quick and secure AI model deployment. Integrated into the Nebius AI Cloud, these microservices come with NVIDIA Enterprise licensing, offering a cloud-native set of tools that enhance AI performance and security. NVIDIA NIM simplifies the deployment of AI models by eliminating the need for custom infrastructure, thereby accelerating biological research and drug discovery processes. Users can deploy these microservices with minimal effort, selecting their preferred configurations and accessing them through standard APIs. The NVIDIA Blueprint for virtual screening in drug discovery further streamlines scientific workflows by integrating multiple microservices into a cohesive, containerized process that facilitates protein folding, molecule generation, and docking. This integrated approach not only speeds up research but also reduces the complexities and costs associated with AI deployment in scientific environments, enabling teams to focus on their core scientific goals without being hindered by technical challenges.

Nebius, a purpose-built AI cloud provider, addresses the ongoing scarcity of AI compute resources by enhancing its platform with new features aimed at improving elasticity and transparency. The introduction of Capacity Blocks and Capacity Dashboard allows organizations to better manage GPU resources, providing clear visibility into available capacity and facilitating efficient planning for machine learning teams. Capacity Blocks offer an intuitive overview of reserved GPU capacity and help align resources with workload needs, while the Capacity Dashboard and API offer real-time transparency and predictability in GPU availability across Nebius AI Cloud. These tools aim to streamline capacity management, allowing teams to make informed decisions, optimize their use of infrastructure, and maintain productivity without disruptions, as they prepare for a gradual rollout in early 2026. Nebius commits to continuously evolving its platform to empower AI practitioners with the compute resources necessary to achieve their goals effectively.

Nebius AI Cloud 3.1 aims to enhance AI workload performance while maintaining transparency and ease of use, focusing on key improvements such as deploying NVIDIA Blackwell Ultra systems in Europe, optimizing network operations, and introducing efficient capacity management tools like Capacity Blocks and Dashboards. The platform also enhances developer experience with updates to Object Storage, JavaScript SDK, and MLOps integration, supporting teams with improved orchestration and monitoring capabilities. Security and governance are bolstered through precise access controls and enhanced Identity and Access Management, while healthcare and life sciences teams benefit from expanded observability and AI-driven tools for scientific research. Overall, Nebius continues to evolve its AI Cloud to offer a powerful, developer-friendly infrastructure that aligns with industry needs and regulatory requirements.

Nebius Token Factory provides a comprehensive architecture for managing the complex infrastructure needed to support large-scale conversational AI platforms, focusing on optimizing latency, cost, and quality. The guide details how the architecture, designed specifically for chat traffic, integrates model shaping, inference system design, and workload-specific engineering to handle the unpredictable and emotional nature of chat interactions. By emphasizing the importance of quantization, sampler correctness, cache efficiency, and speculative decoding, the system ensures consistent user engagement and economic efficiency. With a focus on compliance and security, Token Factory allows teams to operate chat models without compromising governance standards, offering zero-retention inference and regional data isolation. By applying lessons learned from a proof of concept with a major conversational AI platform, Nebius Token Factory enables organizations to deploy chat systems that are stable, predictable, and aligned with real user behavior, without the need for extensive in-house R&D teams.

Nebius Token Factory offers an advanced post-training platform that transforms large open-source models into optimized, production-ready assets without requiring a dedicated ML infrastructure team. By integrating techniques like LoRA, quantization, and reinforcement fine-tuning, the platform allows teams to leverage user data for continuous model optimization and customization, essential for bridging the gap between base models and real-world applications. Nebius Papyrax, a multi-node distributed training framework, enhances scalability, resilience, and efficiency, providing a comprehensive solution for handling large-scale language model training. The platform's robust architecture supports a range of customization and optimization features, such as speculative decoding and multi-precision computation, enabling enterprise-grade performance within secure, private environments. This empowers organizations like Cosine to achieve performance parity with proprietary models while maintaining data sovereignty and security standards.

Nebius has announced a $3 billion agreement to deliver AI infrastructure to Meta over five years, which, along with a deal with Microsoft, will enhance their AI Cloud by 2026, expecting over 2.5 GW of contracted power by then. Their participation in the MLPerf® Training v5.1 benchmark demonstrated strong performance on NVIDIA Blackwell systems, showcasing their commitment to high standards in AI model training. In partnership with Anyscale, Nebius aims to improve platform integration for Ray-based workloads. Recent updates include enhanced support structures, improved management tools in IAM for project oversight, and a focus on security with encryption basics. New documentation and guides cover a range of technical aspects, from managing file access with Soperator to invoking Nebius AI Cloud API methods using grpcurl, as well as integrating Object Storage with Terraform. Nebius has also been active in major industry events, such as KubeCon + CloudNativeCon and Slush, highlighting their AI Cloud capabilities.

Saturn Cloud is a platform that automates the infrastructure management necessary for AI teams to focus on their work from day one, bypassing the extensive setup typically required with raw Kubernetes. By managing GPU node pools, Python environments, and secure data access, Saturn Cloud allows infrastructure engineers to concentrate on strategic initiatives while providing AI teams with ready-to-use, scalable environments. It installs into a Managed Kubernetes cluster in Nebius AI Cloud, offering both managed and self-managed deployment options. The installation involves using a Helm chart and Terraform recipes to create necessary node pools and microservices, such as Fluent-bit for logging and Prometheus for monitoring. Upon setup, users can run development workspaces and jobs, which support various IDEs and allow for the deployment of machine learning models and dashboards. Saturn Cloud resources are portable and can be built from public container images or Docker, integrating seamlessly with CI/CD systems. The platform's resources are free for approved Nebius accounts, ensuring accessibility for organizations seeking efficient AI/ML resource management.