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
Alibaba released open weights for Qwen3.8-2.4T-A95B, its largest open-weight model at 2.4T total parameters with 95B activated per token, and this covers serving it with configurable reasoning on GB300 NVL72.
NVIDIA Dynamo separates vision encoding from prefill and decode in multimodal serving, citing up to 5x faster first-token latency and 7x faster end-to-end response for image-heavy, short-output workloads, plus a 42% cut to text-request latency in mixed traffic.
Meta's Muse Glimmer, a 30B open-weight dense model with a 120K-plus context window, running local agentic workflows across NVIDIA edge, desktop and workstation platforms.
Agents that review a bug report, implement and test a fix, push a patch and ping a human are useful precisely because they act without supervision, which is also the risk. Four ways to deploy them more securely.
An agentic pipeline in NVIDIA's TileGym repo automatically translates cuTile Python and Triton-TileIR GPU kernels into cuTile Rust, porting all 24 public operators at 99.5% of the original Python performance with machine-checked verification.
NVIDIA's SWE-Serve benchmark evaluates coding agents on 53 real SGLang inference-engineering tasks, finding patches that pass conventional tests fail live-serving checks about a third of the time, with pass@1 across 11 models ranging from 35% to 75.5%.
NVIDIA details how NIM's autotuned kernels, tensor parallelism, prefix and state reuse, and speculative decoding lift Nemotron 3 Ultra serving throughput, with a way to benchmark the trade-offs against your own traffic.
Agentic systems trade accuracy against cost, with the best proprietary models and harnesses expensive to run. A harness profile that closes some of that gap for a smaller open model.
Long training runs lose goodput every time a GPU disappears. Nonuniform tensor parallelism adapts the TP degree in place, overlapping the resharding so the cost stays under 1%, and boosts clocks on surviving GPUs so data-parallel replicas do not fall out of sync.
Reinforcement learning has moved from RLHF for assistants to verifiable-reward workflows for reasoning and agent tasks, and this covers the techniques that make it a practical tool rather than a research one.
Benchmarking 45 agent-generated simulation pipelines found that prompt specificity changed code structure and reusability but not physical correctness, with agent skills and reference files bridging prompts to GPU simulation code.
NVIDIA benchmarks Vera Rubin NVL72 on SemiAnalysis AgentX — replayed coding-agent sessions with real context growth, tool calls and sub-agent spawning — claiming up to 30x throughput per megawatt over GB300 NVL72, credited to MoE runtimes, MXFP4/MXFP8 and session-aware serving.
NVIDIA's supply-chain command center pairs a cuOpt allocation solver with Nemotron 3.5 Lightning post-trained on captured planner decisions via NeMo Anonymizer, Data Designer, and AutoModel; the specialized 30B model beat the larger Nemotron 3 Ultra by 31 points on allocation accuracy.
NVIDIA released TensorRT Model Connect, reference implementations that turn a Hugging Face checkpoint into a deployment bundle with a Python CLI, then load and run it from C++ with no Python at inference time. Custom GPU kernels plug in through TVM FFI; releases ship nightly.
NVIDIA's first-party writeup of TensorRT 11.0 multi-device inference landing in Dynamo-Triton 26.07: one KIND_MODEL instance can now own multiple GPUs behind a single gRPC endpoint, cutting Cosmos video-gen latency from 156.6s to 34.2s across 8 GPUs.
NVIDIA releases Nemotron 3.5 Super VL, an open 120B MoE (12B active) multimodal model for agentic workflows. The post covers speculative decoding, NVFP4 quantization, NeMo customization, and reported partner post-training gains, all vendor-reported.
Building an agent does not end at picking a model, because strengths and cost profiles shift between workloads and even within one. NVIDIA NeMo Switchyard routes each step of an agentic task to the model that suits it.
A three tier method for measuring whether an agent skill actually helps: static checks, a distinctiveness pass, then live task runs in isolated environments, benchmarked across more than three hundred verified skills.
NVIDIA open-sources Cluster Readiness Engine, a Kubernetes controller that runs real distributed workloads across GPU node groups to certify cluster health, using adaptive fault isolation to narrow failures down to specific nodes.
NVIDIA details the TensorRT 11.1 Global Performance Tuner: it expands over 200 compiler knobs into build routes, filters by accuracy, picks the fastest engine, supports resumable tuning, and reports 1.14x to 1.63x speedups on ten workloads.
NVIDIA engineers explain how TensorRT Model Connect was designed for coding agents: parallelizable work, per-model-family isolation, reversible changes, GPU-backed validation, and self-improving tests, with humans owning acceptance criteria.
NVIDIA details NVLink 6's layered resiliency: silicon-level FEC and retry, credit-based flow control, NMX failover in seconds, and Dynamo's Shadow Engine Recovery cutting inference restore time from 283s to 7.3s.
NVIDIA's PAIR beta is a virtual router, not an inference engine: it discovers local machines, tracks engine and model readiness, and dispatches independent requests to whichever node is free. Ollama and LM Studio interfaces stay unchanged, so agent harnesses need no edits.
The harness is the architecture around the model, and how it renders context, executes actions, manages state and decides a task is done shapes outcomes as much as the model does. Six capabilities that separate good harnesses from bad.