Articles

NVIDIA says a hosted RL loop lifted Nemotron 3 Nano from 22% to 91% accuracy on a math task for under $5, ending with a downloadable LoRA adapter.

NVIDIA says ModelExpress reduced DeepSeek-V4 Pro startup from 8 minutes to 1 minute 44 seconds by moving weights directly over GPU-to-GPU RDMA.

AI infrastructure competition is being measured in training throughput, not just chip availability. NVIDIA says Blackwell Ultra reached 1,648 TFLOPs per GPU on DeepSeek-V3 671B, about 3x prior delivered performance.

Bristol Myers Squibb is adding a second DGX SuperPOD built on eight DGX Vera Rubin NVL72 systems. The move turns AI infrastructure from a specialist resource into a shared platform for researchers across the company’s global drug-discovery pipeline.

NVIDIA’s SIGGRAPH update shifts physical AI from cloud demos toward edge deployment. The package includes the 4B Cosmos 3 Edge world model, a Synthetic Video Detector NIM microservice, and a DGX Station agent stack built around Nemotron 3 Ultra.

NVIDIA says Nemotron 3 Embed now leads LMEB, with the 8B model ranked first and the 1B model second. The linked Hugging Face discussion cites LMEB scores of 64.4 for 8B and 61.5 for 1B BF16, extending the release beyond its earlier RTEB win.

Retrieval models are becoming a direct quality and cost lever for RAG and agents. NVIDIA says Nemotron 3 Embed 8B ranks first overall on RTEB, with 32k context and smaller 1B variants.

Video analytics development is moving from hand-built pipeline wiring toward natural-language instructions plus coding agents. DeepStream 9.1 adds 13 agentic skills and JetPack 7.2 support.

NVIDIA’s new Jetson T3000 and T2000 modules push Blackwell-class edge AI into smaller robotics systems. The real shift is cost and deployment: 865 FP4 TFLOPS, Cosmos 3 Edge, and memory-saving agent skills are being packaged for robots that cannot depend on cloud inference.

NVIDIA showed Cosmos 3 Nano rising from 54.41% zero-shot accuracy to 93.35% after LoRA and TAO AutoML on a traffic safety video QA task. The result frames agent-run post-training as a practical physical AI workflow.

NVIDIA’s first AI Model Co-Design post argues that LLM dimensions can be as important as scale for inference performance. Its rules of thumb include 128-aligned dimensions, preference for 256 or 512, NVFP4, and parallelism strategies for MoE models.

NVIDIA detailed Vera, a CPU designed for agentic AI workloads where tool calls, code execution, retrieval, and verification sit between model calls. The company claims 50% higher IPC than Grace and 1.8x sustained per-core performance versus x86 on agentic execution workloads.