Blog

Training speculative decoding draft heads at scale introduces two bottlenecks: the memory cost of the language-modeling loss and the attention pattern used by EAGLE-style draft heads. In this article, we 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.

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

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

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

Custom Speculator Training is now in Token Factory. Move from production data to a workload-specific draft model, then deploy it alongside the base model, in one workflow.

Introducing Aether 3.6. Our latest quarterly platform release responds to how AI teams now operate at scale and with maturity, delivering a smoother developer experience, a stronger security and compliance foundation, advances across our in-house storage portfolio, and natural-language control over your AI cloud.

Introducing Nebius Echo, an AI agent built right into the Nebius console. Ask questions, inspect resources, and create infrastructure in plain language. No setup required; find it right inside the console.

MLPerf® Training 6.0 results are in. Across six configurations on NVIDIA Blackwell Ultra systems, Nebius posted the #1 single-node HGX B300 times for Llama-3.1-8B and GPT-OSS 20B pre-training and came within 3.1% of the fastest GB300 NVL72 result on all three benchmarks at 72 GPUs. This post covers the full results and methodology.

Nebius AI Cloud Logs now stream into Datadog Log Management. If you already run on Datadog, you can investigate your Nebius workloads alongside the rest of your stack and trace an AI incident across every layer without switching tools.

Using Nebius Agents Blueprint, we built a regulatory compliance audit agent and ran the same audit task through four configurations — from a $470 prototype to a production system at 72% lower cost with 1.00 recall. This post covers what changed at each step, what each configuration revealed about the one before it, and the production failures that only became visible once the operational architecture existed to find them.

Today we’re introducing the Nebius Agents Blueprint, an open reference architecture for building, operating, and continuously improving AI agents in production. This post covers the six-component composable stack — inference, orchestration, retrieval, grounding, observability, and simulation — and the case study behind it: a compliance audit agent that achieved 72% lower cost and 20% higher precision over a GPT-based prototype by improving the system, not the model.

Nebius has collaborated with NVIDIA to bring two retail AI blueprints to production on Nebius infrastructure: the NVIDIA Agentic Commerce Blueprint and the NVIDIA Retail Catalog Enrichment Blueprint for automated product content. Both are open-source reference architectures built on NVIDIA NIMs that retailers and developers can customize and deploy today with a 1-click deployment on Nebius AI Cloud.