NVIDIA-NeMo
16 Tools · 93 Intel
Tools
NVIDIA's Python API for defining large language models and running optimized inference on NVIDIA GPUs.
World foundation models, datasets, and tooling packaged together for physical AI and robotics work.
Real-time full-duplex speech-to-speech model with persona control via text role prompts and voice conditioning.
Reference architecture for GPU-accelerated vision agents that search, summarize, and answer questions about video.
NVIDIA foundation model for generalist humanoid robots, with perception-to-action policies that transfer across bodies.
NVIDIA's security scanner that inspects AI agent skills for vulnerabilities and malicious patterns.
AI weather and climate prediction workflows get built and deployed through this open-source deep-learning framework.
Lightweight C++ inference engine that runs LLMs and vision-language models on edge devices.
Text normalization and inverse text normalization toolkit built for ASR and TTS pipelines.
Collection of logits processors that customize LLM generation, working with vLLM, Hugging Face, and TensorRT-LLM.
Rust proxy that routes LLM traffic across providers and translates OpenAI and Anthropic APIs.
NVIDIA GEAR Lab initiative applying world models to the robotics data problem in embodied AI training.
Intel
Training runs out of GPU memory long before it runs out of compute. Host offloading in JAX parks selected activations in pinned host memory and streams them back for the backward pass, beating rematerialization by up to 57% on large MoE workloads — but only when the copies overlap real work.
NVIDIA details a memory-driven Chief of Staff agent that keeps derived knowledge in Markdown pages separate from a SQLite obligation ledger, uses an intent gate to prioritize stated goals over urgency, and exposes an inspectable, correctable preference policy, improving benchmark accuracy from 82.8%
Siemens EDA used NVIDIA NeMo Gym verifiers to find failure modes in a chip-design agent, then tuned skills and MCP tool descriptions. Benchmark score rose from about 55% to 95% while turns dropped from 155 to 53.
NVIDIA details its Open Agent Safety Platform, pairing the open-source OpenShell sandboxed runtime on Vera CPUs with Sentry monitoring on BlueField-4 DPUs to enforce policy and observe agent behavior out-of-band at line speed.
Generative recommenders replace embedding based retrieval with sequence modeling and transformer style architectures such as HSTU and semantic IDs, which changes how cold start and the long tail behave. The companion repository carries production implementations and fused kernels.
Self-hosting a coding assistant in a regulated or source-sensitive environment runs into three problems: source cannot leave the network, the assistant invents package names that create supply-chain risk, and outputs need validating. NeMo Guardrails addresses them.
NVIDIA's Earth-2 team shows how Score-Based Data Assimilation constrains diffusion weather models (StormCast, CorrDiff) with live observations, cutting wind-speed RMSE up to 54%, plus a transformer-based HealDA model for fast global state estimates.
NVIDIA reports its AVO agent architecture lifting a frontier model from a 30 percent baseline to a perfect score on ARC-AGI-3, crediting persistent memory, supervision and tool use rather than the model alone.
NVIDIA walks through the Rubin GPU as a machine for sustained multi-step inference rather than single prompts: HBM4 capacity and bandwidth that keep large KV caches resident, a third-generation Transformer Engine at NVFP4 precision, NVLink 6 scale-up, and rack-scale power smoothing in the NVL72.
NVIDIA walks through using agents on a local RTX Spark hub to plan, build assets for, and train a Microduck robot in Isaac Sim and Isaac Lab, with 4,096 parallel environments and local policy validation before sim-to-real.
A walkthrough for post-training a four billion parameter world model into a manipulation policy small enough to run on an onboard robotics module, then evaluating it in closed loop simulation.
NVIDIA walks through adapting its pretrained X-Mobility policy into robot-specific residual specialists via RL, orchestrated by an agent workflow with human approval gates at each stage. Evaluation compares base and residual policies under identical seeds and goals; runtime publishes velocity comman
NVIDIA's AIPerf replaces GenAI-Perf with a multiprocess client architecture so the benchmarking tool itself doesn't bottleneck high-concurrency LLM load tests, supporting 15+ endpoint types and configurable, realistic arrival patterns.
NVIDIA explains how FLARE 2.8/2.9 separate persistent federation services from job workers, adding Docker, Kubernetes, and Slurm support so federated learning can run across mixed HPC and cloud infrastructure.
NVIDIA walks through using an AI coding agent with a custom 'skill' to migrate a ROS 2 node to CUDA-backed zero-copy buffers, then verifies the result with Nsight Systems traces.
NVIDIA walks through open C++ samples that pair ONNX Runtime with the TensorRT RTX provider for local inference. They cover ASR models, SAM 2.1 and FLUX.2-klein, with DGX Spark throughput numbers and a quantization workflow that leaves application code unchanged.
NVIDIA safety and security teams map the agent stack into models, harnesses, meta-harnesses, secure runtimes and inference infrastructure, and argue that controls belong at the layers an agent cannot rewrite.
NVIDIA and Jane Street engineers walk through implementing Mamba-2's forward pass in cuTile with three custom kernels, using Tensor Core matrix multiplies, latency hints and fusion to keep intermediates in registers.
A federated training design for multimodal models that exchanges only small adapter weights over a frozen backbone, cutting per-client traffic from tens of gigabytes a round to a fraction of one.
NVIDIA engineers detail a central identity-gateway pattern: one platform session validated by stateless regional gateways against a shared Redis session store, replacing per-cluster OIDC flows across federated Kubernetes and AI platforms.
NVIDIA open-sources the Run:ai platform: KAI Scheduler, a GPU Fractioning Operator, quota management and a workload contract. Partners including Nebius, Crusoe and Mirantis are integrating it for GPU orchestration on Kubernetes.
NVIDIA profiled 163,000 agentic sessions and found 97% have unique execution shapes, which makes multi-design CPU fleet strategies impractical. Vera is pitched at the mixed load — per-thread speed on the latency-bound sequential path, concurrency for fan-out bursts — with a first-party 1.5x per-core
NVIDIA's TensorRT Edge-LLM ran Qwen3.6-27B on a single Jetson AGX Thor at 52 tokens/sec, finishing MLPerf's Edge Agentic benchmark 6.4x faster than llama.cpp using NVFP4 quantization, FP8 KV cache and tree-based multi-token prediction.
NVIDIA lays out a framework for agent evaluation that scores both intermediate tool-call steps and final task-completion state, argues for executable environment checks over LLM-as-judge, and benchmarks Nemotron 3.5 Lightning at 86% on PinchBench.