The auto-updated technology brief
NVIDIA introduces DFlash speculative decoding to boost inference performance by up to 15x on Blackwell GPUs, improving low-latency AI workflows by optimizing autoregressive LLM token generation.
NVIDIA demonstrates an open source example combining Hermes Agent with NVIDIA NemoClaw to accelerate and secure product research across Outlook, Slack, and GitHub, leveraging AI agents for data synthesis and decision-making.
NVIDIA explores Vision-Language-Action (VLA) and World-Action Models (WAM) that leverage pretrained large-scale VLM backbones to enable robots to generate actions from visual and language inputs.
Mixture-of-experts (MoE) models enable larger AI model capacity by activating only a subset of parameters per token. NVIDIA explores advanced fusion kernels to improve MoE training throughput for large-scale AI systems.
NVIDIA introduces a method to bridge the gap between training and deployment of autonomous vehicle models using closed-loop post-training with their Alpamayo platform, improving vision-language-action models for complex driving scenes.
NVIDIA leads in agentic coding performance on the industry's first multi-vendor AI agent benchmark, Artificial Analysis AgentPerf, which profiles real-world AI coding tasks.
Foundation models pretrained on large protein and genomic sequence datasets, such as ESM2 and Evo 2, enable diverse computational biology tasks. NVIDIA BioNeMo recipes facilitate efficient fine-tuning of these models using LoRA.
Transformer architectures power many large language and generative AI models. As these models grow, training demands increase GPU resources and engineering time. Optimizing transformers for low-precision training accelerates development and reduces costs.
NVIDIA introduces MiniMax M3, a multimodal AI system on NVIDIA accelerated infrastructure including NVIDIA Blackwell, enabling long-context reasoning and simplifying enterprise AI workflows.
NVIDIA introduces Nemotron 3 Ultra, enhancing efficiency and reasoning capabilities for long-running AI agents managing complex workflows with growing token counts.
AI agents are evolving beyond chat to perform complex tasks in enterprises, requiring secure and governed environments to protect sensitive data and ensure safe operation.
Amazon Bedrock's new Fully Managed Knowledge Bases simplifies building enterprise RAG pipelines by providing native data connectors, Smart Parsing for automatic multi-format data preparation, and an Agentic Retriever for complex multi-step queries—all integrated with AgentCore Gateway so developers can focus on business outcomes rather than infrastructure management.