Nebius submitted results across seven system configurations to&nbsp;MLPerf&reg; Inference v6.1, our broadest round to&nbsp;date. We&nbsp;took five first-place results, including both DeepSeek R1 scenarios on&nbsp;the full-rack Nebius GB300 NVL72, and were one of&nbsp;only two providers to&nbsp;post preview results on&nbsp;NVIDIA Vera Rubin NVL72, where we&nbsp;led per-GPU Server throughput on&nbsp;DeepSeek R1.

MLPerf&reg; Inference v6.1: Previewing NVIDIA Vera Rubin NVL72 and leading full-rack Nebius system with NVIDIA GB300 NVL72 results

Today we&rsquo;re launching the Nebius AI&nbsp;Builder Program: one place to&nbsp;build and scale AI&nbsp;systems on&nbsp;open models and independent tools, with no&nbsp;lock-in to&nbsp;a&nbsp;single vendor. Members get $400+ in&nbsp;credits and discounts across the stack, cookbooks with working code, free courses from Nebius Academy and NVIDIA DLI, certifications, expert office hours, and a&nbsp;community of&nbsp;builders.

Introducing the Nebius AI&nbsp;Builder Program: build AI&nbsp;systems you own on&nbsp;an&nbsp;open, independent stack

The MLPerf&reg; Storage v3.0 results are&nbsp;in. Nebius Object Storage fed 768&nbsp;simulated NVIDIA B200 accelerators on&nbsp;the RetinaNet training workload, the highest accelerator count of&nbsp;any submission this round, and ran the highest Unet3D accelerator count among object storage submissions. This post covers the full results, methodology, and what they mean for training directly on&nbsp;object storage.

MLPerf&reg; Storage v3.0 results: Leading RetinaNet training on&nbsp;Nebius Object Storage

A&nbsp;detailed analysis of&nbsp;the incident on&nbsp;August 19, 2026&nbsp;that led to&nbsp;service outages in&nbsp;the us-central1 region. The incident was caused by&nbsp;a&nbsp;storm-related event at&nbsp;the data center facility that disabled the cooling infrastructure.

Incident post-mortem analysis: us-central1 service disruption on&nbsp;August 19, 2026

Nebius is&nbsp;he&nbsp;first AI&nbsp;cloud to&nbsp;adopt NVIDIA Groq 3&nbsp;LPX, to&nbsp;Nebius Token Factory, adding generation-optimized performance purpose-built for agentic&nbsp;AI. Independently benchmarked at&nbsp;3,400 output tokens per second for a&nbsp;single user on&nbsp;Google Gemma 4 31B, NVIDIA Groq 3&nbsp;LPX pairs with NVIDIA Vera Rubin NVL72 so&nbsp;developers get it&nbsp;through the same Token Factory catalog and API they already use.

Nebius Token Factory Becomes First AI&nbsp;Cloud to&nbsp;Adopt NVIDIA Groq 3&nbsp;LPX

The results are in&nbsp;for the Nebius Serverless AI&nbsp;Builders Challenge. In&nbsp;total, 35&nbsp;projects qualified across AI&nbsp;&amp; ML, Scientific AI&nbsp;&amp; Healthcare, and robotics, each with a&nbsp;public repo and a&nbsp;technical blog post. Discover the top three winners, special award recipients, and cast your vote for the Community Choice award.

Serverless AI&nbsp;Builders Challenge: winners announced

Nebius Academy is&nbsp;the education and research hub of&nbsp;Nebius. This guide lays out everything it&nbsp;offers in&nbsp;one place: AI&nbsp;adoption programs and readiness assessments for organizations, role-based AI&nbsp;cloud certifications with free credits, open courses for professionals, and research grants that fund academic work with GPU compute.

Nebius Academy: closing the gap between AI&nbsp;tools and the people who use them

Nebius was built on&nbsp;the conviction that AI&nbsp;must stay competitive and diverse. Open source is&nbsp;essential to&nbsp;keeping it&nbsp;that way.

The whole AI&nbsp;stack moves faster in&nbsp;the open

Training speculative decoding draft heads at&nbsp;scale introduces two bottlenecks: the memory cost of&nbsp;the language-modeling loss and the attention pattern used by&nbsp;EAGLE-style draft heads. In&nbsp;this article, we&nbsp;show how Streaming Cross Entropy removes the vocabulary-logits bottleneck and block-sparse FlashAttention addresses the draft-head attention pattern, improving memory and performance efficiency for long-context training.

Training speculative decoders: removing the logits and attention bottlenecks

Nebius and PyTorch built a&nbsp;high-performance DeepSeek V3 training recipe on&nbsp;TorchTitan, running on&nbsp;256&nbsp;Nebius B200 GPUs with Soperator. This post dives deeper into why GPU-initiated RDMA with NVSHMEM and DeepEP make such an&nbsp;impact in&nbsp;improving performance for large MoE models, and walks through the full IBGDA and DeepEP setup you can run on&nbsp;Nebius today.

Inside the Nebius + PyTorch DeepSeek V3 recipe: NVSHMEM and DeepEP for wide expert parallelism

SlimSpec is&nbsp;a&nbsp;low-rank draft LM-head architecture for speculative decoding that reduces projection cost while preserving full-vocabulary support. In&nbsp;this article, we&nbsp;examine why the draft LM&nbsp;head becomes a&nbsp;bottleneck, how SlimSpec differs from vocabulary-truncation methods and what our EAGLE-3 experiments show about end-to-end decoding performance.

SlimSpec: faster speculative decoding without cutting the vocabulary

LangChain has released a&nbsp;Deep Agents profile tuned for NVIDIA Nemotron 3&nbsp;Ultra, achieving top agent accuracy among open models at&nbsp;roughly 10x lower cost than closed alternatives&nbsp;&mdash; with the model itself untouched. The Nebius Agents Blueprint already pairs Deep Agents with Nemotron 3&nbsp;Ultra on&nbsp;Nebius Token Factory, and support for the tuned profile is&nbsp;coming soon.

Blog