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

Q2 2025: Nebius AI&nbsp;Cloud updates

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

Agent 101: Launching&nbsp;production-grade agents at&nbsp;scale

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

Nebius monthly digest: June 2025

We&nbsp;kicked off Q2 with a&nbsp;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&nbsp;Studio has been updated significantly.

Nebius AI&nbsp;Studio Q2&nbsp;2025&nbsp;updates

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

From genome analysis to&nbsp;quantum chemistry: Nebius powers the next generation of&nbsp;biotech research with NVIDIA

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

Introducing Managed Soperator: Your quick access to&nbsp;Slurm training

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

Kubernetes: How to&nbsp;use it&nbsp;for AI&nbsp;workloads

Our AI&nbsp;R&amp;D team announces the open-source release of&nbsp;the SWE-rebench dataset of&nbsp;more than 21,000 real-world, interactive software engineering tasks. For a&nbsp;detailed methodology and technical report, please see our accompanying <a href="https://arxiv.org/abs/2505.20411">paper on&nbsp;arXiv</a>.

SWE-rebench dataset: More than 21,000 verifiable tasks for SWE agents

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

The role of&nbsp;compute cluster networking for AI&nbsp;training and inference

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

Nebius monthly digest: May 2025

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

Nebius demonstrates industry-leading AI&nbsp;training performance in&nbsp;latest MLPerf&reg; results

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

Introduction to&nbsp;model distillation: Efficient knowledge transfer for AI&nbsp;applications

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

Incident post-mortem analysis: outage of&nbsp;the S3 service in&nbsp;the eu-north1 region

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

Improving AI&nbsp;cluster observability: New metrics, Grafana dashboards and advanced logging

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

SWE-rebench: A&nbsp;continuously updated benchmark for&nbsp;SWE&nbsp;LLMs

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

Serving Qwen3 models on&nbsp;Nebius AI&nbsp;Cloud by&nbsp;using SkyPilot and SGLang

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

Serving Llama 4&nbsp;models on&nbsp;Nebius AI&nbsp;Cloud with SkyPilot and SGLang

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

Nebius monthly digest: April 2025

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

Understanding the Model Context Protocol: Architecture

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

Reasoning critics enable better parallel search for software engineering agents

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

What is&nbsp;Apache Spark and how can it&nbsp;help with LLMs?

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

Bulk Object Storage data migration with SkyPilot

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

Meet Audit Logs in&nbsp;Nebius&nbsp;AI&nbsp;Cloud

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

Understanding pre-trained AI&nbsp;models and their applications

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

AI&nbsp;model fine-tuning: What it&nbsp;is&nbsp;and why it&nbsp;matters

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

Nebius AI&nbsp;Studio Q1&nbsp;2025&nbsp;roundup: Fine-tuning, new models and major expansions

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

Storing personal data in&nbsp;Nebius&nbsp;AI&nbsp;Cloud: Security, GDPR and AI&nbsp;readiness

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

Announcing the integration between Nebius and dstack

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

Supercharging startups: Nebius&nbsp;accelerates AI-native innovation with NVIDIA

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

Q1 2025: Nebius&nbsp;AI&nbsp;Cloud updates

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

Watch the talks: Videos from Nebius AI&nbsp;Cloud Unveiled meetup

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

Nebius monthly digest: March 2025&nbsp;and NVIDIA&nbsp;GTC

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

Nebius AI&nbsp;Cloud is&nbsp;now integrated with SkyPilot

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

Meet Managed Service for MLflow in&nbsp;general availability

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

JupyterLab&reg;: The first app&nbsp;in&nbsp;our new service for&nbsp;AI&nbsp;development

In&nbsp;this blog post, we&rsquo;ll demonstrate how to&nbsp;fine-tune an&nbsp;LLM using Nebius AI&nbsp;Studio with a&nbsp;function-calling task as&nbsp;our running example. The goal isn&rsquo;t to&nbsp;achieve state-of-the-art results but to&nbsp;walk you through the steps and showcase why fine-tuning matters.

Beyond prompting: Fine-tuning LLMs with Nebius AI&nbsp;Studio

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

GPU vs&nbsp;CPU: what is&nbsp;the best bioinformatics accelerator?

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

Incident post-mortem analysis: networking issues for Managed Services for Kubernetes on&nbsp;March 13, 2025

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

Nebius and Outerbounds form strategic technology partnership through integration

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

Nebius monthly digest, February&nbsp;2025

Today marks a&nbsp;significant step forward for AI&nbsp;developers and businesses: fine-tuning is&nbsp;<a href="https://nebius.com/services/studio-fine-tuning?utm_medium=blog&amp;utm_source=blog&amp;utm_campaign=ai-studio-fine-tuning">now available</a> in&nbsp;Nebius AI&nbsp;Studio. This launch transforms how you can customize leading open-source models, making them work precisely for your specific needs.

Make AI&nbsp;work for you: fine-tuning launches on&nbsp;Nebius AI&nbsp;Studio

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

We&rsquo;re introducing the 300&nbsp;MW&nbsp;New Jersey region and&nbsp;expanding to&nbsp;Iceland

<a href="/services/managed-postgresql">Managed Service for PostgreSQL</a> is&nbsp;now generally available to&nbsp;all Nebius users. We&rsquo;ve been improving the service&rsquo;s functionality and stability for the last few months, to&nbsp;empower AI&nbsp;practitioners with a&nbsp;reliable and convenient tool for storing structured data in&nbsp;the cloud.

Introducing general availability of&nbsp;Managed Service for PostgreSQL

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

What are AI&nbsp;compute clusters and how to&nbsp;choose yours?

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

What is&nbsp;cloud infrastructure and why&nbsp;is&nbsp;it&nbsp;essential for&nbsp;AI&nbsp;development

Images with pre-installed drivers in&nbsp;our Managed Service for Kubernetes enhance NVIDIA GPU cluster scalability and reduce compute node start-up time by&nbsp;2-3x.
Preparing a&nbsp;GPU cluster requires some effort and careful attention from your engineering team. It&nbsp;can become a&nbsp;significant challenge in&nbsp;terms of&nbsp;deploying and configuring a&nbsp;compute environment. Today, we&rsquo;re happy to&nbsp;announce a&nbsp;small yet meaningful enhancement that makes the cluster preparation process much easier and faster.

Pre-installed drivers are now available for&nbsp;Kubernetes nodes

Today, we&rsquo;re <a href="https://github.com/nebius/kvax">open-sourcing Kvax</a>, our FlashAttention implementation based on&nbsp;JAX. Designed for efficient training with long sequences, Kvax supports context parallelism and optimized computation of&nbsp;document masks. It&nbsp;outperforms many other FlashAttention implementations in&nbsp;long-context training with dense packing, achieving state-of-the-art performance.

Kvax: Fast and easy-to-use FlashAttention implementation for JAX

A&nbsp;wave of&nbsp;advancements, including the generation of&nbsp;massive amounts of&nbsp;data, is&nbsp;transforming how researchers and companies reshape science. Access to&nbsp;compute and workflow management systems is&nbsp;essential to&nbsp;benefit from the new wave of&nbsp;tools. In&nbsp;this blog post, we&rsquo;ll explore how to&nbsp;leverage <a href="https://seqera.io/platform/">Seqera Platform</a> and Slurm on&nbsp;Nebius AI&nbsp;Cloud to&nbsp;deploy scalable <a href="https://www.nextflow.io/">Nextflow</a> workflows.

Running Nextflow workflows with Seqera Platform and Slurm on&nbsp;Nebius AI&nbsp;Cloud

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

How we&nbsp;streamlined HR&nbsp;operations with an&nbsp;AI&nbsp;assistant (and why you should too)

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

What is&nbsp;a&nbsp;token in&nbsp;AI? Understanding how AI&nbsp;processes language with tokenization

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

Nebius monthly digest, January&nbsp;2025

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

Incident post-mortem analysis: Major connectivity loss on&nbsp;January 27, 2025

Striving to&nbsp;simplify the user experience, we&nbsp;<a href="https://console.eu.nebius.com/">increased</a> the amount of&nbsp;compute resources immediately available for every user that signs up&nbsp;to&nbsp;our cloud. <a href="https://console.eu.nebius.com/">You can now</a> access 8 <a href="https://www.nvidia.com/en-us/data-center/h200/">NVIDIA H200 GPUs</a>, 16 <a href="https://www.nvidia.com/en-us/data-center/h100/">NVIDIA H100 GPUs</a> and 2 <a href="https://www.nvidia.com/en-us/data-center/l40s/">NVIDIA L40S GPUs</a> via our console and start using them immediately.

Nebius extends immediate access to&nbsp;NVIDIA Hopper GPUs

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

Creating images with Flux: Your prompt guide

If&nbsp;you think New Year celebrations started on&nbsp;January 1, it&nbsp;depends on&nbsp;how you look at&nbsp;it&nbsp;&mdash; because in&nbsp;the AI&nbsp;world, the real fireworks went off in&nbsp;December with the release of&nbsp;<a href="https://studio.nebius.ai/playground?models=deepseek-ai%2FDeepSeek-V3">DeepSeek V3</a>. Now, <a href="https://studio.nebius.ai/playground?models=deepseek-ai%2FDeepSeek-R1">DeepSeek R1</a> is&nbsp;lighting up&nbsp;the sky with open-source brilliance that&rsquo;s making even the most entrenched Silicon Valley giants feel the heat.

DeepSeek R1 and V3: Chinese AI&nbsp;New Year started early

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

Meet SDKs for Go&nbsp;and&nbsp;Python

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

Transforming ideas into images: Introducing text-to-image generation on&nbsp;Nebius AI&nbsp;Studio

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

Using SkyPilot and Kubernetes for multi-node fine-tuning of&nbsp;Llama 3.1

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

Orchestrating LLM fine-tuning on&nbsp;K8s with SkyPilot and MLflow

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

Incident post-mortem analysis: Partial unavailability on&nbsp;November 26, 2024

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

A&nbsp;landmark year: highlights of&nbsp;Nebius&rsquo; journey in&nbsp;2024

In&nbsp;this follow-up to&nbsp;our previous research blog post, we&nbsp;focus on&nbsp;the data collection process used to&nbsp;train our action generator and critic models. We&rsquo;re also releasing two datasets on&nbsp;Hugging Face: <a href="https://huggingface.co/datasets/nebius/SWE-bench-extra">nebius/SWE-bench-extra</a>, containing 6,411 Issue/Pull-Request pairs, and <a href="https://huggingface.co/datasets/nebius/swe-agent-trajectories">nebius/SWE-agent-trajectories</a>, featuring 80,036 software engineering agent trajectories, where an&nbsp;agent attempts to&nbsp;solve these issues.

Scaling data collection for&nbsp;training software engineering agents

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

Q4: Nebius cloud updates

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

What is&nbsp;prompt engineering?

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

Nebius AI&nbsp;Studio expands with vision models, new language models, embeddings and LoRA

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

AI&nbsp;model performance metrics: in-depth guide

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

AI&nbsp;model hosting options

We&nbsp;are now <a href="#blackwell-form">accepting pre-orders</a> for <a href="https://www.nvidia.com/en-us/data-center/gb200-nvl72/">NVIDIA GB200 NVL72</a> and NVIDIA HGX B200 clusters to&nbsp;be&nbsp;deployed in&nbsp;our data centers in&nbsp;the United States and Finland from early 2025. Based on&nbsp;<a href="https://www.nvidia.com/en-us/data-center/technologies/blackwell-architecture/">NVIDIA Blackwell</a>, the architecture to&nbsp;power a&nbsp;new industrial revolution of&nbsp;generative AI, these new clusters deliver a&nbsp;massive leap forward over existing solutions.

Nebius opens pre-orders for NVIDIA Blackwell GPU-powered clusters

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

Nebius monthly digest, November&nbsp;2024

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

Creating your own AI-powered code generator and reviewer

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

What is&nbsp;a&nbsp;context window in&nbsp;AI? Understanding its importance in&nbsp;LLMs

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

Building an&nbsp;AI-powered finance planner with full-stack Next.js and Nebius AI&nbsp;Studio

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

PostgreSQL in&nbsp;the context of&nbsp;ML

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

Inference optimization techniques and solutions

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

Nebius opens its first availability zone in&nbsp;the&nbsp;United States

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

Leveraging training and search for better software engineering agents

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

Nebius opens offices in&nbsp;the US

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

Nebius monthly digest, October&nbsp;2024

At&nbsp;Nebius, we&rsquo;re committed to&nbsp;supporting the AI&nbsp;community, which is&nbsp;why we&rsquo;re introducing our Explorer Tier, designed to&nbsp;make GPUs more affordable and accessible for those just getting started. H100 is&nbsp;now priced at&nbsp;$1.5/h for the first 1,000 hours monthly.

Explorer Tier: Investing in&nbsp;AI&nbsp;entrepreneurs

This blog post will guide you through the differences, use cases and cost-benefit analysis of&nbsp;different AI&nbsp;models, especially in&nbsp;the context of&nbsp;inference.

How to&nbsp;choose between large and small AI&nbsp;models: A&nbsp;cost-benefit analysis

Discover how Nebius AI&nbsp;Studio enables you to&nbsp;integrate the famous Llama&nbsp;3.1&nbsp;405B large language model in&nbsp;your applications.

How to&nbsp;run Meta Llama&nbsp;3.1 405B with Nebius AI&nbsp;Studio API

We&nbsp;have developed a&nbsp;new version of&nbsp;Nebius platform that we&nbsp;believe will serve your needs even better. It&nbsp;is&nbsp;already being tested by&nbsp;our in-house LLM R&amp;D team and a&nbsp;number of&nbsp;clients. Now, we&nbsp;are rolling it&nbsp;out to&nbsp;everyone.

We&rsquo;re launching a&nbsp;new Nebius platform built from the ground&nbsp;up

The current expansion phase will enable us&nbsp;to&nbsp;place up&nbsp;to&nbsp;60,000 GPUs at&nbsp;the M&auml;nts&auml;l&auml; location alone.

Nebius to&nbsp;triple capacity at&nbsp;Finland data center to&nbsp;75&nbsp;MW

In&nbsp;September, we&nbsp;announced the opening of&nbsp;our French region, launched the world&rsquo;s first open-source K8s operator for Slurm as&nbsp;well as&nbsp;Nebius AI&nbsp;Studio, a&nbsp;product for GenAI builders&nbsp;&mdash; and there&rsquo;s more.

Nebius monthly digest, September&nbsp;2024

Our new platform not only allows application builders to&nbsp;use generative AI&nbsp;without significant effort but does so&nbsp;at&nbsp;a&nbsp;fraction of&nbsp;the cost. Our goal is&nbsp;to&nbsp;democratize access to&nbsp;cutting-edge AI&nbsp;technologies, enabling businesses of&nbsp;all sizes to&nbsp;innovate and compete in&nbsp;the AI&nbsp;space.

Introducing Nebius AI&nbsp;Studio: Achieve fast, flexible inference&nbsp;today

Our Paris availability zone will be&nbsp;one of&nbsp;the first in&nbsp;Europe to&nbsp;offer <a href="https://www.nvidia.com/en-us/data-center/h200/">NVIDIA H200 Tensor Core GPUs</a>. This is&nbsp;just the beginning as&nbsp;we&nbsp;continue to&nbsp;expand our operational capacity to&nbsp;meet fast-growing demand for high-quality infrastructure and the latest GPUs from AI&nbsp;builders around the world.

Nebius launches its first availability zone in&nbsp;France

We&rsquo;ve built an&nbsp;open-source Kubernetes operator that runs and manages Slurm clusters as&nbsp;Kubernetes resources. From this article, you&rsquo;ll find out how the community tackled this task previously and the details of&nbsp;the architecture of&nbsp;our solution, which we&nbsp;named Soperator.

Explaining Soperator, Nebius&rsquo; open-source Kubernetes operator for Slurm

We&rsquo;ve developed Soperator, a&nbsp;Slurm-based workload manager with a&nbsp;modern and simplified user experience for ML&nbsp;and HPC cluster orchestration.

Introducing Soperator: a&nbsp;Kubernetes Operator for&nbsp;Slurm

Machine learning inference tests a&nbsp;model against real-world unseen data. Understand what is&nbsp;inference, and how models are deployed in&nbsp;production.

What is&nbsp;inference?

One of&nbsp;the experts from our LLM R&amp;D team, Alexander Golubev, recently took part in&nbsp;the Kaggle competition hosted by&nbsp;the Large Model Systems Organization (LMSYS). Here are his impressions.

Chatbot Arena Human Preference Predictions: tech&nbsp;review

Explore our comprehensive guide to&nbsp;automated machine learning (AutoML) and find out why it&rsquo;s changing data science and making machine learning available to&nbsp;everyone.

What is&nbsp;AutoML? Understanding automated machine learning

Spark has recently been gaining popularity over Hadoop for ML/AI use cases. Is&nbsp;Hadoop becoming obsolete? Do&nbsp;these technologies compete or&nbsp;complement each other? This blog post answers all your questions about Spark and Hadoop.

Spark&nbsp;vs. Hadoop in&nbsp;data engineering

Features are essential to&nbsp;understanding data patterns and training machine learning models. Learn about their importance and generation techniques.

What is&nbsp;a&nbsp;feature in&nbsp;machine learning?

Entropy is&nbsp;a&nbsp;machine learning term borrowed from thermodynamics that measures randomness or&nbsp;disorder in&nbsp;any system. Why measure disorder? Mathematics uses entropy to&nbsp;measure this chaos&nbsp;&mdash; or, more specifically, the probability of&nbsp;chaos.

Entropy in&nbsp;machine learning&nbsp;&mdash; applications, examples, alternatives

We&nbsp;recently reduced the prices of&nbsp;NVIDIA H100, invited everyone interested to&nbsp;reserve H200, launched a&nbsp;public preview of&nbsp;Managed Service for MLflow and released a&nbsp;new talk about deploying generative AI&nbsp;models in&nbsp;production.

Nebius monthly digest, August&nbsp;2024

MLflow is&nbsp;a&nbsp;renowned industry tool that streamlines workflows in&nbsp;the model development cycle. We&nbsp;made MLflow more accessible to&nbsp;a&nbsp;broad audience of&nbsp;ML&nbsp;enthusiasts by&nbsp;providing it&nbsp;as&nbsp;a&nbsp;managed solution.

Announcing Managed Service for MLflow in&nbsp;public preview

In&nbsp;this post, we&rsquo;re comparing open-source and proprietary software, with a&nbsp;detailed focus on&nbsp;definitions, differences, key advantages and business models.

Open-source vs&nbsp;proprietary software

By&nbsp;uncovering surprising links between seemingly unrelated LLM architectures, a&nbsp;way might be&nbsp;paved for effective idea exchange and boosting efficiency.

How transformers, RNNs and&nbsp;SSMs are more alike than you think

Drawing from Nebius&rsquo; and our clients&rsquo; extensive experience, this guide and research aims to&nbsp;help engineers choose the most fitting storage solutions for deep learning.

Choosing storage for deep learning: a&nbsp;comprehensive guide

Machine learning experiments help you discover the most optimum model version for your specific use case. Read this article to&nbsp;learn about different types of&nbsp;experiments and what you need to&nbsp;watch out for when conducting them.

AI research and development at Nebius

Our blog

Projects

Kvax

SWE-rebench

Join our team

ML Engineer, LLM Training and Inference Optimization

Senior/Staff ML Engineer, AI Research

Senior Software Engineer, Nebius AI Studio Fine-Tuning

Senior Software Engineer, LLMs and Agentic Systems

Start your journey today

Explore Nebius

Products

Resources

Solutions

Prices

Programs

Company

Legal