Blog

We introduce LK losses — training objectives that directly optimize the acceptance rate for speculative decoding draft models. They are a drop-in replacement for KL divergence, with no computational overhead, and they work with any draft architecture and any target model size, to deliver consistent improvements in inference throughput across models ranging from 8B to 685B parameters. We are open-sourcing our trained draft models (LK-Speculators) and training datasets (Infinity-Instruct-Completions). An implementation of LK losses is also available as a pull request to SpecForge.

The results of our MLPerf® Inference v6.0 submission demonstrate Nebius’ ability to maximize efficiency for modern AI inference workloads on the latest NVIDIA Blackwell and Blackwell Ultra platforms.

Physical AI is moving from controlled demos to real-world deployment — and that shift demands benchmarks grounded in actual operations, not lab conditions. Today, Nebius joins Positronic as a founding consortium partner of the Physical AI Leaderboard (PhAIL), a new platform that evaluates robot AI models on real hardware using commercially relevant tasks and production-grade metrics.

The Nebius VPN Gateway CLI provides a simple way to configure and operate site-to-site VPN connectivity in Nebius AI Cloud. In this post, we walk through how it enables infrastructure-as-code workflows for IPSec gateways, helping teams manage secure and reliable connectivity across cloud and on-prem environments.

The serverless services at Nebius are a natural extension of how an AI infrastructure cloud evolves over time, building on a mature and well-established underlying platform. As the platform develops, it becomes possible to expose compute in more flexible and elastic forms that better match how AI workloads are consumed.

NVIDIA RTX PRO 6000 Blackwell opens new opportunities for cost-efficient inference and increased performance for visual computing and scientific simulations.

In collaboration with PyTorch, Nebius helped demonstrate up to 41% faster pre-training of DeepSeek-V3 models on NVIDIA Blackwell GPUs.

A detailed analysis of the incident on March 10, 2026 that led to service outages in the us-central1 region.

At NVIDIA GTC 2026, Nebius and DataRobot, with NVIDIA, introduced a validated AI Factory stack for production-grade agent workloads. In this post, we outline how the DataRobot Agent Workforce Platform runs on Nebius AI Cloud to support sustained inference, governance and cost control for AI agents deployed in live business workflows.

Nebius and Eigen AI are partnering to bring optimized frontier open-source models to Nebius Token Factory. As part of the collaboration, optimized implementations of models such as DeepSeek, GLM, GPT-OSS, Kimi, Llama, MiniMax and Qwen will be published on the platform, giving developers direct access to high-performance inference through production-ready endpoints and APIs.

Nexla and Nebius are partnering to deliver a production-ready data and agent stack that connects governed enterprise data with infrastructure built for sustained inference. In this post, we outline how this architecture enables multi-agent systems to move from fragmented data pipelines to reliable production workflows, and show it in action through a live “Inspiration to Trip” demo presented with Tripadvisor at NVIDIA GTC.

A detailed analysis of the incident on February 26, 2026 that led to service outages in the eu-north-1 region.

Today we’re introducing the Inference Frontier Program, a new builder-to-builder initiative dedicated to production inference systems. The program surfaces real architectures, optimizations and engineering tradeoffs from teams running large-scale inference in production.

Modern ML and LLM workloads require environments equipped with specialized hardware, high-performance networking and integrated MLOps tools. In this article, we’ll explore how AI-focused clouds differ from general-purpose platforms — and what criteria define the right provider for building scalable AI systems.

NVIDIA Nemotron 3 Super is now available on Nebius Token Factory, bringing a 120B hybrid MoE model optimized for multi-agent systems and complex reasoning workflows to production deployments. With long-context inference and OpenAI-compatible APIs, teams can run Nemotron 3 Super through dedicated GPU endpoints and autoscaling infrastructure without managing their own serving stack.

Self-hosted AI agents offer control and flexibility, but they also introduce real security risks. Incidents involving malicious ClawHub skills, exposed default ports and prompt-injection attacks show that running OpenClaw is not just an installation task, but an infrastructure decision. This guide explains OpenClaw’s architecture and maps real threats to concrete-hardening controls, so teams can deploy it safely in production.

We’re introducing SWE-rebench-V2, the next iteration of our large-scale dataset of reinforcement learning (RL) environments for training autonomous software engineering agents (SWEs).

Today, Nebius and Toloka are announcing plans to bring Tendem into the Nebius ecosystem. This integration further strengthens the Nebius AI stack, anchoring the raw intelligence of Token Factory and the autonomy of Tavily agentic search with a programmable layer of human reliability. Originally designed as the market’s pioneer hybrid human-AI agent, Tendem is now the first platform to embed vetted human experts directly into agentic workflows — making expert judgment callable via the Model Context Protocol (MCP), the emerging standard for AI tool integration.

We are introducing Dedicated Endpoints and a Custom Weights Hub in Nebius Token Factory. You can now choose GPU type, define GPUs per replica, set scaling limits, select region and deploy your own model weights to isolated endpoints. Deployment becomes a defined, controllable part of your production architecture.

NVIDIA and Nebius ran joint benchmarks using NVIDIA Run:ai, the AI workload orchestration and optimization software platform on the Nebius AI Cloud. The goal was simple: test whether fractional GPU allocation could improve efficiency and scalability for real-world inference workloads — without compromising performance.

When inference becomes distributed, routing strategy can determine whether a system scales or stalls. In this article, we examine a real agent-style workload where cache-aware routing in vLLM reduced average step time by nearly 50 percent and cut P95 latency from over a minute to under 20 seconds — with the same model, hardware and traffic.

As AI adoption accelerates, energy use increasingly sets the boundaries of how far the systems can scale. Power availability, efficiency and infrastructure design are becoming practical constraints. This shift is prompting to think of concrete ways to manage the energy footprint of AI systems by optimizing energy use and creating measurable efficiency gains. In our latest whitepaper, we explain how Nebius improves efficiency across the stack, from software engineering to hardware design and data center operations.

We recently introduced support for exporting billing data from Nebius AI Cloud in the FOCUS format. This update is a small but important step toward making cloud integration simpler and financial operations smoother for teams building and scaling AI workloads. At Nebius, we believe billing data should be easy to work with, easy to integrate and easy to trust — especially when AI infrastructure costs are a core part of your business model.

Large mixture-of-experts language models promise significant gains in model quality, but deploying them in real products often exposes hard latency limits. This article explains why MoE systems that look efficient in benchmarks struggle under production constraints, and how architectural decisions around routing, batching and serving determine whether latency budgets hold under worst-case inputs and real user behavior.

While reinforcement learning drives agents to state-of-the-art performance, rejection fine-tuning serves as a powerful baseline. Stemming from our extensive experiments with different models and scaffoldings, we are sharing a dataset of 67k high-quality OpenHands trajectories from Qwen3 Coder 480B for research purposes. We also include two RFT checkpoints — Qwen3 Instruct 30B and 235B, achieving ~50% and ~60% on SWE-bench Verified respectively.

Bioscience seeks to unravel the structure of protein complexes that drive nearly every cellular process. Open-source tools and data banks are essential for accelerating advances in this field. With support from Nebius AI Cloud and a Nebius Academy fellowship, researchers at the University of California, San Francisco are adapting widely used protein modeling software to modern GPU cloud platforms, expanding access and impact for the scientific community.

This guide teaches the essentials of serving large language models with vLLM. It builds from foundational neural network concepts, like transformers and attention, to introduce practical inference workflows, explore useful vLLM features and offer hands-on, operational guidance for efficient deployments.

In this blog, we will take a look at how teams can deploy the four NVIDIA NIMs available now in our platform, how they can be consumed via their API and how they can fit into larger scientific workflows. It also covers the Virtual Screening Blueprint, which turns folding, molecule generation and docking into a ready workflow that becomes available in JupyterHub shortly after deployment.

The Nebius AI Cloud 3.1 release continues the evolution of our platform into a transparent, orchestrated, developer-first environment — where performance is accessible, capacity is visible and control feels effortless. Across new platforms based on NVIDIA Blackwell Ultra, smarter capacity management and an improved experience for both developers and platform operators, every enhancement moves us closer to that goal.

Europe’s first operational NVIDIA GB300 NVL72 systems are now live in Nebius AI Cloud. The Blackwell Ultra-based platform is fully deployed in our Finland data center, delivering supercomputer-class performance for large-scale training and high-throughput inference. This rollout expands the compute foundation available to teams building next-generation models across Europe.

Today, we’re expanding Nebius AI Cloud’s elasticity even further with the introduction of Capacity Blocks and Capacity Dashboard — two new capabilities that bring additional clarity to GPU capacity management and give users full confidence in the compute they rely on.

This guide reveals the engineering patterns behind modern conversational platforms, and how Nebius Token Factory shapes open models, serving pipelines and GPU economics to keep quality stable at scale. The blueprint presented here is a proof of concept (PoC) grounded in work we conducted with a major conversational AI platform operating hundreds of thousands of concurrent chats. The details are anonymized, but the architectural lessons are real.

Post-training by Token Factory is now live, the missing layer between MVP and real production AI. Base models are great for demos, but production needs models trained on your data that are faster, cheaper, and aligned to your workflows. Today, Token Factory makes that possible for every team, without building ML infrastructure.

November at Nebius brought yet another multi-billion hyperscaler deal and further expansion of our data center footprint, new leading benchmark results and deeper platform ecosystem integrations.

Saturn Cloud is an AI development platform that is deployed from your Nebius console. It enables AI/ML teams to build and deploy models with their favorite tools, without reliance on engineering support. You can use it for free and watch your AI/ML team onboard and start shipping right away.

We’re proud to share the results of our participation in the MLPerf® Training v5.1 benchmark, where Nebius showcased strong performance across several configurations of the latest NVIDIA Blackwell and Blackwell Ultra systems. This round continues our commitment to transparency and collaboration with the MLCommons® community, as we work to ensure the highest quality standards for training and fine-tuning next-generation GenAI models.

Launching biomolecular models at scale in a structured and safe environment is key for drug discovery teams to reach reliable, experimentally relevant insights faster. This article provides a practical, reproducible blueprint for running Boltz-2 inference in Nebius, from single-GPU experiments to scalable multi-node screening pipelines.

Nebius and Anyscale have partnered to offer deeper platform integration and enable teams to deploy and scale Python and AI workloads with Ray easier, faster and cheaper. The partnership combines Nebius AI Cloud and Anyscale’s managed Ray platform, delivering a complete stack for developers and platform teams to move multimodal AI — across data processing (structured and unstructured), training and serving — from code to production.

Research on SWE agents involves building and running thousands of containers, quickly surpassing the limits of a single host. Our AI R&D team unveils the large-scale infrastructure that powers this research — the backbone behind recent releases such as SWE-rebench — and an open-source part of it to support the broader community.

This autumn has been one of acceleration for Nebius — from the rollout of the latest AI Cloud platform release to our brand new next-gen UK data center and the launch of Nebius Token Factory, an enterprise-grade inference platform.

We are introducing updates to the Nebius Status Board to enhance transparency during service outages and provide users with a clearer view of Nebius AI Cloud service availability.

We’re pleased to announce that Nebius AI Studio now hosts NVIDIA Nemotron Nano 2 VL, a compact, production-ready multimodal reasoning model engineered for real-world document intelligence and video understanding.

Let’s explore one of the key features that makes the new NVIDIA GB200 NVL72 stand out: the fifth generation NVIDIA NVLink™ scale-up fabric. We’ll discuss how it redefines infrastructure by moving beyond the traditional 8-GPU NVLink. You’ll see a practical example of how to take advantage of this capability. Finally, we’ll examine a real-world use case: pre-training the Nemotron-4 340B LLM.

We’re launching Managed SkyPilot API Server on Nebius AI Cloud. It’s a fully managed service that transforms SkyPilot from a single-user tool into a shared platform where teams can pool resources, coordinate workloads and stop worrying about infrastructure operations.

At Nebius, we’ve spent the past year working closely with enterprises that are moving AI projects from experiments to business-critical systems. The challenges they raise aren’t about “getting more GPUs” — they’re about how to govern, secure and scale AI infrastructure without creating bottlenecks for their teams. That’s the backdrop for our latest AI Cloud 3.0 release, named “Aether.”

Today, we are thrilled to announce we achieved major security and compliance milestones. Independent third-party audits have verified that our security controls meet the requirements of SOC 2 Type II (including HIPAA), and align with the principles of NIS2 and DORA. We also obtained ISO 27001 certification, strengthened our practices by incorporating principles from ISO 27701, 27018, 27799, 27032, and standalone ISO 22301.

Serving AI companies and enterprises with text-to-video inference is no small feat. These teams demand enterprise-ready performance — at scale, with low latency, and high reliability. In this post, we’ll unpack the state-of-the-art engineering that enables Nebius and Baseten to deliver production-grade video generation — and show you how to test it yourself.

September was a landmark month for Nebius. From a major new customer for our AI infrastructure to industry-leading performance recognition, we’ve made strides that directly strengthen the systems you rely on.

We’re proud to announce that Nebius is one of the first NVIDIA Cloud Partners to achieve NVIDIA Exemplar Status on NVIDIA H200 GPUs for training workloads. This recognition validates that Nebius meets NVIDIA’s rigorous standards for performance, resiliency, and scalability — addressing one of the most pressing challenges in AI infrastructure: ensuring consistent workload performance and predictable cost across clouds.

Before AI can generate text, answer questions or summarize information, it first needs to read and understand human language. That’s where tokenization comes in. A tokenizer takes raw text and breaks it into smaller pieces or tokens. These tokens may represent whole words, parts of words or even individual characters and each is mapped to a unique numerical ID that models can process mathematically. In this article we’ll explore how tokenizers work, examine common approaches and walk through the basics of building one yourself.

This guide walks you through building a production-ready, multi-agent AI system by using the Google ADK and A2A, powered by Nebius AI Studio models. With sentiment detection, RAG-powered answers and escalation handling, you can automate customer queries end-to-end.

Model distillation is a practical way to shrink large models into efficient versions that run faster and cost less. As parameter counts climb into the billions, model distillation LLM makes it possible to cut GPU memory use, speed inference and simplify deployment. In this blog we’ll explain how the method works, why GPU compute matters, and what to keep in mind when moving from research models to production systems.

Learn how to build a smart, scalable content generator by using Nebius AI’s Llama 3.3-70B and Qdrant’s vector search. This RAG-based system lets you upload brand-specific documents and get custom social posts, article drafts and more, rooted in your actual company data.

A detailed analysis of the incident on September 3, 2025 that led to service outages in the us-central1 region.

The incident impacted API operations and Console functionality due to persistent routing loops between network domains, while other regions remained operational.

Jupyter Notebook is a browser-based tool for interactive coding, data exploration and documentation. It lets you run code step by step while combining results, visualizations and explanations in one place. Widely used in machine learning, it speeds up experimentation, ensures reproducibility and makes collaboration easier. This article looks at how Jupyter supports ML workflows, its key features and the tasks it handles best.

Today, we’re happy to share our new performance milestone — the latest submission of MLPerf® Inference v5.1 benchmarks, where Nebius achieved leading performance results for three AI systems accelerated by the most in-demand NVIDIA systems on the market: NVIDIA GB200 NVL72, HGX B200 and HGX H200.

Choosing between a single node and a cluster is one of the core infrastructure decisions when working with LLMs. The choice directly affects training speed, resource efficiency and operational costs. In this article we’ll explain how single-node and cluster configurations differ, when each works best and what to consider before choosing one.

In August, we introduced self-service NVIDIA Blackwell GPUs in Nebius AI Cloud and published several in-depth technical articles, including ones on cluster reliability and liquid cooling. We also continued to cover customer success — all this and more in the latest digest.

A Virtual Private Cloud (VPC) is a cornerstone of modern cloud infrastructure. It gives organizations the ability to isolate resources, control traffic and configure security much like a private data center — while keeping the flexibility of the cloud. In this article, we’ll look at what a VPC is, how it’s built and why it has become the standard environment for running applications and machine learning workloads.

Open-source, global-scale infrastructure is crucial to accelerate scientific innovation. Developed in the ELIXIR BioHackathon and enabled by Nebius, the BioHackCloud project is designed as a reference implementation for a standards-based, federated multi-cloud platform — having successfully demonstrated an Attested TLS proof of concept for private LLM inference and confidential task execution.

AI models need massive computing power, and GPUs have become the backbone for training and inference. This article explains what GPU servers are, why they matter for AI and how teams can access GPU compute through cloud platforms, dedicated instances, bare-metal servers or hybrid setups. It also covers how to choose the right approach based on workload type.

When starting a job, you expect it to run without interruptions. This expectation holds true across many domains, but it resonates especially deeply with machine learning engineers who launch large-scale pre-training jobs. Maintaining a stable training environment is crucial for delivering AI results on schedule and within budget constraints.

Modern language models are great at generating text: they complete phrases, answer questions and write code. But in scenarios that demand analytical thinking or step-by-step problem solving, producing a likely next word isn’t enough. What’s needed is a logically consistent answer that unfolds in a structured way. That’s where reasoning models come in.

This guide explains what an epoch means in day-to-day ML pipelines and workflows, how to recognize when you’ve trained enough, and which factors determine a sensible stopping point.

Cluster orchestration tools automatically coordinate the distribution of AI workloads across thousands of clusters. They can scale clusters up or down to match usage and handle failure conditions without interrupting operations. This article explores various functions and types of cluster orchestration tools, including best practices that promote efficiency.

NVIDIA HGX B200 instances are now publicly available as self-service AI clusters in Nebius AI Cloud. This means anyone can access NVIDIA Blackwell — the latest generation of NVIDIA’s accelerated computing platform — with just a few clicks and a credit card.

When you’re training a machine learning model, you’ll often hear terms like epochs, iterations and batches. They’re sometimes used interchangeably, but each one refers to something different. Knowing the difference helps you train your model more effectively.

Our July’s digest describes how our customers like Stanford and Shopify use flexible compute capacities and the steps we are taking to boost performance. Nebius’ first anniversary also took place in July, marked by Nasdaq in Times Square, and there was news from across the ocean as well.

In this article, we break down the inference pipeline — from weight loading and batching logic to routing, autoscaling and monitoring — and explore how engineering choices at each level affect end-to-end behavior in production.

Slurm Workload Manager is a cornerstone of high-performance computing (HPC) infrastructure, trusted by supercomputing centers worldwide for its scalability and flexibility. As AI workloads grow in size and complexity, Slurm is gaining traction among ML teams as well. In this article, we will look at why it remains relevant, how it supports GPU clusters and what to consider when using it in AI workflows.

AI training and inference are critical stages in the machine learning lifecycle, each with distinct objectives and computational requirements. Let’s explore the key distinctions between AI training and inference, their unique resource demands and best practices for building scalable ML workflows.

Learn what few-shot learning is, how it compares to traditional machine learning approaches and why it’s transforming how we train AI with minimal data.

Training and running large AI models can be powerful, but they often come with heavy costs in speed, memory and infrastructure. Model distillation helps cut through this complexity. It works by transferring knowledge from a large “teacher” model into a smaller “student” model, allowing teams to maintain strong performance while making deployment faster, lighter and more affordable. In this article, we’ll explore why distillation matters, how compute makes it practical at scale and what it means for building efficient AI systems.

Skip the CLI commands and web console clicks — just ask Claude about your cloud infrastructure. Today, we’re excited to announce the Nebius MCP Server, our integration that connects Claude by Anthropic, or other AI chatbots, to the Nebius AI Cloud infrastructure.

DeepSeek-V3 is an open-weight model built specifically for engineering use. It can be deployed locally, fine-tuned and adapted to your team’s infrastructure. In this article, we break down its architecture, behavior and practical applications — from code generation to systems that rely on logical reasoning.

Welcome back to our quarterly roundup featuring all the improvements and updates delivered on Nebius AI Cloud over the past three months. This overview highlights the key developments we’ve made to enhance your AI infrastructure experience.

To go from prototype to production, AI agents need more than just a good model. In this guide, we break down the four components that matter most: reliable LLMs, orchestration frameworks, evaluation tools, and memory systems. We cover how teams are using Nebius AI Studio with CrewAI, ADK, LangChain, and more to ship scalable, observability-friendly agent workflows, all powered by fast, cost-efficient inference.

Our standout moment in June was NVIDIA GTC Paris, where Nebius was recognized multiple times. To mark the conference, we announced the opening of the UK cloud region and introduced integration with NVIDIA DGX Cloud Lepton, NVIDIA AI Enterprise stack and other systems. Also, our ISEG2 is #13 on the Top500 of supercomputers; we held the AI Discovery Award, and launched Managed Soperator for Slurm-based workloads. Read these and other news in today’s digest.

Choosing the optimal number of epochs for training a machine learning model is about finding balance. Too few epochs lead to underfitting: the model performs poorly on both training data and unseen examples. Too many epochs cause overfitting: the model achieves high accuracy on the training set, but struggles with real-world tasks. In this article, we’ll explore the main factors that influence how many epochs to use — from dataset size and model complexity to early stopping techniques.

We kicked off Q2 with a single mission: turn raw compute horsepower into concrete business outcomes. With numerous launches, including groundbreaking models, streamlined fine-tuning, scalable throughput and seamless integrations, Nebius AI Studio has been updated significantly.

As part of NVIDIA GTC Paris at VivaTech, Nebius has announced deeper integration of the NVIDIA AI Enterprise software suite. This includes NVIDIA BioNeMo, a collection of tools, applications, generative AI solutions, and pre-trained microservices (NVIDIA NIM) designed squarely for the biopharma sector.

Learn the fundamentals of object storage, how it differs from traditional block storage solutions and why it is becoming the go-to choice for modern data management. With data volumes exploding and cloud systems becoming the norm, traditional systems are struggling to keep up with today’s data chaos. That’s where object storage comes in — built for the cloud, made to scale and ready for anything. This guide cuts through the jargon to show you why object storage is the future and how it outperforms block storage where it counts.

Training machine learning models can cost anywhere from tens of thousands to millions of dollars depending on model size, dataset volume and infrastructure. In this article we look at the main cost components of training in the cloud, what drives the final bill and how to optimize spending without compromising results.

Managed Soperator, our fully managed Slurm-on-Kubernetes solution, is now available for everyone in self-service. It provides you with a ready-to-work Slurm training cluster, powered by modern hardware and delivered with all necessary pre-installed libraries and drivers — allowing you to start ML training immediately.

Building and deploying AI systems at scale means juggling complex infrastructure — and that’s where Kubernetes shines. From managing GPU resources to scaling inference endpoints, Kubernetes brings structure and automation to the chaos of machine learning pipelines. In this article, we’ll break down how Kubernetes works, why it’s a natural fit for AI workloads and what best practices help keep things resilient, reproducible and production-ready.

While earlier machine learning models could be trained on CPU servers with one or two GPUs, today’s generative AI models have billions of parameters — orders of magnitude more than their predecessors. Such models require terabytes of training data that can only be parallel processed over multiple GPU servers. These GPU servers work together in clusters to run the underlying data computations that make these models work. This article explores GPU cluster networking technologies and their critical role in accelerating AI workloads.

Even larger workloads are welcome! We raised quotas last month, so you can now access up to 32 NVIDIA H200 GPUs on demand in both the US and Europe. Speaking of Europe, we’ll be revealing many of our plans for the region at NVIDIA GTC Paris next week.

Today, we’re thrilled to announce our first submission of MLPerf® Training v5.0 results. As a peer-reviewed industry benchmark suite, MLPerf® Training by MLCommons® is one of the most trustworthy sources of data about AI cloud performance in the industry.

Model distillation is a powerful technique in machine learning where a compact “student” model learns to replicate the behavior of a larger, more complex “teacher” model within the given task. In this tutorial, we demonstrate how to perform distillation by using Nebius AI Studio, to create a grammar-correcting model.

A detailed analysis of the incident on May 5th, 2025 that led to an outage of the S3 service in the eu-north1 region.

In the past weeks, we have made several significant observability-related improvements to the Nebius AI Cloud. Among them: advanced monitoring metrics in our web console and API, out-of-the-box Grafana dashboards, as well as monitoring and logging features, allowing customers to upload their custom metrics and logs to our cloud.

Our AI R&D team presents SWE-rebench, a new benchmark for evaluating agentic LLMs on a continuously updated and decontaminated set of real-world software engineering tasks mined from real GitHub repositories. Our goal with this benchmark is to make evaluation of software engineering LLMs more transparent, reproducible and focused on core model capabilities.

Explore how to get Qwen3 running on Nebius AI Cloud with SkyPilot and SGLang. This setup enables you to deploy both the massive 235B MoE model and the efficient 32B variant seamlessly, leveraging high throughput, cost-effective scaling and robust multilingual support.

Let’s walk through how to get Llama 4 running on Nebius AI Cloud (recently integrated with SkyPilot) by using SGLang as the serving framework. This combo provides high throughput, efficient memory usage and none of the typical deployment headaches.

This month, we revealed some of the partnerships and collaborations we’ve been engaged in, such as with Meta and DDN. We also showcased work done by our customers like vLLM and Brave, as well as participants of our Research credits program — including Stanford, USC and others.

In this short research blog post, we discuss how reinforcement learning-finetuned reasoning language models can be used as a better alternative to regression-based critics when performing parallel trajectory search in test time, and how to train such models to prioritize precision over recall.

As LLM-powered agents become more complex, integrating them with tools, APIs, and private data sources remains a major challenge. Model Context Protocol (MCP) offers a clean, open standard for connecting language models to real-world systems through a modular, plug-and-play interface. In this article, we explore how MCP works.

Large language models (LLMs) rely on fast data processing and distributed computing, making the efficiency of data processing tools a critical factor. Apache Spark streamlines text data preparation, enables parallel processing of massive datasets and simplifies the development of scalable ML workflows. This article explores Spark’s architecture, its advantages for data preparation and solutions to common limitations when working with LLM-scale models.

Moving large datasets between S3 buckets is often slow, unreliable and frustrating — especially across accounts or clouds. In this post, we share a fast, fully open-source workaround using SkyPilot, s5cmd and Nebius AI Cloud.

Audit Logs is a new feature, now available in preview in Nebius AI Cloud. With Audit Logs, our customers can improve their accountability, security and ensure compliance with regulatory requirements.

Fine-tuning pre-trained AI models is a challenging task in itself. What strategy should you use and what resources do you require? This article presents a comprehensive overview of the AI model fine-tuning process.

Pre-trained models let you skip the hassle of training new models by providing already hosted models you can fine-tune to build use-case-specific models. This article explores how this works, some use cases, upsides and downsides (plus countermeasures) of using pre-trained models.

We’ve had a very productive first quarter of 2025 at Nebius AI Studio, shipping major features, adding powerful new models and significantly expanding our infrastructure. This comprehensive roundup covers everything we’ve launched and where you can find us next.

The increasing adoption of AI cloud solutions highlights the need for efficient, vendor-independent orchestration tools. dstack provides an open-source, AI-focused orchestration platform that emphasizes flexibility and ease of use for AI development. We are announcing an integration with dstack to enhance the developer experience for professional ML teams. You can now choose Nebius in dstack and start managing dev environments, executing training jobs and deploying models on our AI infrastructure.

If you work with personal data and need to store it in the cloud, you have to choose a cloud provider that offers appropriate storage. If you’re wondering whether our cloud can store and process personal user data, the answer is yes. Our platform ensures the secure storage of personal data, in full compliance with applicable regulations. But what does that mean in practice? Let’s break it down.

Nebius AI Cloud is collaborating with NVIDIA Inception, a leading program for AI startups with over 22,000 members, to accelerate AI startup innovation. Nebius’ AI Lift program — announced at NVIDIA GTC 2025 — offers eligible NVIDIA Inception members up to $150,000 in cloud credits, access to AI technical expertise and support and a range of other exclusive benefits.

The end of the quarter is a great time to reflect and share a kind of changelog with you, highlighting what’s improved in Nebius AI Cloud over the past three months. This quarter, we launched new regions, released new services for MLOps and enhanced our existing product lineup to minimize your infrastructure concerns even further.

We wrapped up our first series of technical meetups with stops in Paris, London and San Francisco — each time providing a deep dive into our cloud. Here, you can watch all the talks from the first event in the series.

NVIDIA GTC was the highlight of last month for us. Our extensive participation included being a platinum sponsor, a strong presence with the booth, two tech talks, interviews with our leaders, partners and customers, and much more.

We’re excited to announce our integration with SkyPilot, an open-source framework that simplifies running AI and batch workloads across cloud platforms. This collaboration enables direct access to Nebius AI Cloud resources via SkyPilot.

Managed Service for MLflow becomes generally available — which means it gains more stability, a newer version of the platform, service level agreements and a service fee.

Today, we’re excited to announce that JupyterLab is now available in our Standalone Applications service, running on NVIDIA GPUs on our cloud platform. This fully managed offering eliminates infrastructure complexities, allowing you to focus entirely on your AI and machine learning (ML) projects.

In this blog post, we’ll demonstrate how to fine-tune an LLM using Nebius AI Studio with a function-calling task as our running example. The goal isn’t to achieve state-of-the-art results but to walk you through the steps and showcase why fine-tuning matters.

Speeding up computational tasks is critical to push the boundaries of science and ensure scalability and efficiency. Check out how NVIDIA Hopper GPUs available on Nebius AI Cloud performed against CPU workflows to accelerate homology research, deep learning-enabled protein embeddings and ML clustering.

On March 13, 2025, a release of one of the control plane components of the VPC service in the eu-west1 region caused massive networking issues for clients' pods of Managed Services for Kubernetes.

We are excited to unveil our latest collaboration: Nebius AI Cloud, the ultimate cloud for AI innovators, and Outerbounds, an AI/ML platform powered by the popular open-source framework Metaflow, are now integrated as part of a broader technology partnership.

In February, we at Nebius were busy preparing several new data center regions and working on our large-scale participation in NVIDIA GTC. We also rolled out some improvements to the platform and contributions to open source — as well as new initiatives to support AI research: a brand new biotech award and a research credits program.

We’re thrilled to announce a major upgrade to our US-based compute capacity. To bring it to life, we’ve joined forces with DataOne, an AI hosting infrastructure company, to ensure that the first phase of the New Jersey facility goes live this summer. We’re also launching a colocation facility in Iceland with Verne, a provider of sustainably powered data centers across the Nordics, and expect it to go live this month.

Generative AI and LLMs require immense computational power, making compute clusters essential for efficient scaling. Training and fine-tuning these models exceed the capacity of a single compute node, and parallel computing enables businesses to handle GenAI workloads at scale. In this article, we explore what compute clusters are and how to choose the right one for your needs.

Images with pre-installed drivers in our Managed Service for Kubernetes enhance NVIDIA GPU cluster scalability and reduce compute node start-up time by 2-3x.

Preparing a GPU cluster requires some effort and careful attention from your engineering team. It can become a significant challenge in terms of deploying and configuring a compute environment. Today, we’re happy to announce a small yet meaningful enhancement that makes the cluster preparation process much easier and faster.

Modern AI systems require substantial computational power to train and infer complex models. Cloud GPUs provide scalable infrastructure with high parallelism, offering access to virtualized GPUs through APIs and cloud platforms. The article delves into the technical aspects of cloud GPUs including virtualization, elastic scaling and management of high-performance workloads.

Repetitive employee queries can overwhelm HR departments and reduce time for strategic tasks. At Nebius, we implemented an AI Assistant powered by Nebius AI Studio that automates over 150 monthly queries. In this article, we explore how NLP and RAG help Studio drive such processes efficiency.

Explore what tokens are, how they enable AI models to understand and process language, represent text in AI systems and why tokenization is critical for LLMs. This post will also allow you to understand different tokenization techniques, how they affect processing efficiency and how modern NLP models like transformers use them to deliver accurate and context-aware responses.

We have been hard at work this past month, expanding the scope of our materials, which support customer progress on the platform. Our blog has been hopping too, showcasing the latest releases from Nebius AI Studio and the resilient infrastructure potential Nebius AI Cloud, accelerated by NVIDIA, brings for different AI use cases. We wish you a productive Q1, as we empower your AI journey on our platform.

A detailed analysis of the incident that led to service outages. The incident impacted Compute API, Virtual Private Cloud and Managed Kubernetes.

With this complete prompt guide, you will learn practical tips for creating exceptional images with Flux on Nebius AI Studio. I will share clear, actionable insights from experience to help bring your AI art game to the next level. With Nebius AI Studio’s prices starting at $0.0013 per image, you can experiment with crystal-clear visuals up to 2000×2000 pixels without breaking the bank.

Today, we are happy to announce that our SDKs for Go and Python are now generally available to all users. This major update in programmatic interfaces brings more flexibility for MLOps teams and AI developers.

From concept art to product visuals, text-to-image AI is revolutionizing creative workflows, reducing hours of design work to mere minutes. Today, we’re excited to launch Nebius AI Studio’s text-to-image service, delivering premium, production-ready image generation at a fraction of typical costs.

When adapting a large language model to your domain or specialized application, you want efficiency and a certain degree of simplicity. The setup of Managed K8s plus SkyPilot we’re going to run through today provides exactly that, albeit not as the only option. Meta Llama-3.1-8B is just an example here — you can apply a similar method to many other LLMs.

While many focus on the modeling aspects during LLM fine-tuning, efficiently managing compute and experiment tracking is equally crucial. This guide demonstrates how to effectively leverage SkyPilot and MLflow, two powerful open-source tools, to orchestrate LLM fine-tuning jobs.

A detailed analysis of the incident that impacted Compute Cloud virtual machines and Managed Kubernetes clusters, including the root cause, mitigation steps, and improvements to prevent future occurrences.

While the holiday season is perhaps not all about numbers — and more about not thinking about them for a while — they still pop up here and there. Whether it’s the GPU capacity you’re planning for next year or how many billions of parameters you’ll use in your next model, numbers are driving us into 2025. Here are some figures we are reflecting on these days — while keeping a sense of fun about them.

After completing a full rewrite of our cloud from the ground up in October, we began improving the foundation that we brought to the market. The features and tools we have introduced since then span our entire range of cloud services, from Compute Cloud with NVIDIA accelerated computing in a new region, to managed MLOps solutions.

As the holiday season approaches and we all juggle year-end projects, we’re excited to bring you a gift of expanded capabilities on Nebius AI Studio. These updates — including advanced vision models, a broader range of language models, powerful embeddings and LoRA hosting — are designed to help you jumpstart the new year with stronger, more versatile AI tools.

Prompt engineering has advanced considerably in the last three years. This article overviews the process and includes several beginner and advanced strategies to help AI teams get the most out of their LLM models.

Metrics provide a mathematical basis to assess AI model quality. Organizations use a range of metrics to measure every aspect of the AI model — from output, to speed, user experience and even ethics. This article covers 30+ metrics you can use to track AI progress.

This blog explores the process of AI model hosting, where trained models are deployed and made accessible via APIs for users and applications. It covers the complexities of running AI workloads in distributed environments, the different technology layers involved and key factors like performance, cost and security when choosing the best hosting option.

This month’s digest is our most packed so far. We’re opening pre-orders for NVIDIA Blackwell GPU-powered clusters, expanding to a new US data center in Q1 2025 and opening offices there. Both of our R&D teams, AI and hardware, shared their latest developments, and more.

In this hands-on tutorial, you’ll learn how to create a powerful code generation and review application that you can customize to your exact needs. We’ll use Nebius AI Studio and open-source models, making it both cost-effective and extensible.

Explore the importance of context windows in AI and LLMs. Learn how context window size affects model performance, tokenization and natural language processing.

In this guide, we’ll build an AI-powered financial planner using the Meta Llama 3.1 70B model via Nebius AI Studio and a full-stack Next.js app. This solution works seamlessly whether you’re using the Pages Router or the App Router, as much of the customization relies on setting up a custom Next.js server.

Among the many database options available, PostgreSQL stands out as a tool of choice for ML engineers. This article explores why and how you can use PostgreSQL to increase efficiency and speed in ML development.

Deep learning models and LLMs work well in development but often slow down on prod due to scale and unpredictability of input. This article explores more than ten different approaches you can use to optimize your model inference.

Scheduled to go live in Q1 2025, the Kansas City availability zone will house thousands of state-of-the-art NVIDIA GPUs, primarily H200 Tensor Core GPUs in the initial phase, with the energy-efficient NVIDIA Blackwell platform expected to arrive in 2025.

In this blog post, we’ll share some of our recent findings on how search can enhance agentic systems for software engineering. Specifically, we’ll demonstrate how we’ve built an SWE-agent-based system that exclusively uses open-weight LLMs and achieves a 40.6% resolved rate on SWE-bench Verified.

Supporting our ambitious growth plans in the US market, we now have dedicated office spaces in San Francisco, Dallas and New York. As the Nebius team expands, our employees in these cities now have a place to hold meetings, spend time with colleagues and take care of growing the business.

October has been so eventful that it’s hard to know where to begin. But the most important news is that you can now take advantage of the newly rebuilt Nebius platform! We also delivered a keynote at TechCrunch Disrupt, reduced prices dramatically, celebrated 1 year since our public launch — and more.