NVIDIA AI
10 Intel
NVIDIA AI is the official X account of NVIDIA, the American technology company known for its graphics processing units and AI computing platforms. The account shares NVIDIA's deep learning and AI research news, product announcements, and developer updates.
Is this you? Sign in with X to claim this profile.
Intel
NVIDIA's Axolotl3D reconstructs occluded 3D geometry from images, camera data, and partial geometry, claiming state-of-the-art results on single- and multi-view 3D generation ahead of ECCV2026.
World Labs' Atlas world model, pretrained on NVIDIA Blackwell GPUs, takes 32 images of a real place and generates new views, reconstructs the scene, and simulates it in Voyager in real time.
NVIDIA and USC's HorizonRelight tackles chunk-based video relighting's visible lighting jumps by propagating context across sliding windows, producing more consistent long-video results with fewer artifacts.
NVIDIA's Nemotron 3 Embed 8B took the #1 spot for combined nDCG@10 on Perplexity's Q2D-Web benchmark, evaluated across 190M web documents and nearly 70K agent-reformulated queries in 10 languages.
NVIDIA details the serving optimizations behind Nemotron 3 Ultra NIM: autotuned kernels, tensor parallelism, prefix and state caching, scheduler tuning and MTP speculative decoding. It reports 1,997 tok/s versus 718 on 4xB200 at 50 TPS/user, with AIPerf steps to reproduce on your own traffic.
NVIDIA releases Nemotron 3.5 Bolt E4B, a Gemma 4 E4B-based open model post-trained for conversational agents. It offers multi-turn tool use, switchable reasoning, 64k context, and measured TTFT and VRAM on RTX and Jetson, plus open recipes and a speculative drafter.
NVIDIA's walkthrough of AIPerf, the GenAI-Perf successor: a multiprocess load client with Poisson and gamma arrivals, trace replay, and percentile TTFT/ITL reporting, demonstrated on vLLM with static and variable-length workloads.
NVIDIA and Siemens EDA describe an evaluation loop for a chip-design agent using NeMo Gym: a gated artifact verifier, rollout trace inspection, and revisions to skills and MCP tool schemas. Over about 25 rollouts, scores rose from roughly 55% to 95% with far fewer turns and tokens.
NVIDIA, Princeton and Maryland researchers find most failed agent rollouts hinge on one early, recoverable mistake. PivotOPD combines preventive and recovery distillation in one PPO update, beating 13 baselines on ALFWorld, WebShop and search QA, with a 3.2 point SWE-Bench Verified gain.
StepFun's Step 5 Preview is a sparse MoE with 600B total and 27B active parameters, 1M context and native tool calling. The NVFP4 checkpoint, vLLM and SGLang recipes, and NeMo fine-tuning paths ship alongside it under StepFun's own license.