AI Lab
Official NVIDIA announcements — model releases, product launches and research, each one summarized with every source covering it, by AIBriefs. RSS
Skild AI's S1 robot foundation model learns unseen, long-horizon tasks from a single video demonstration via in-context learning, with no weight updates or task-specific post-training. Skild has raised nearly $1.7B since 2023 and hit a $100M annual revenue run rate 10 months after first commercial deployment.
NVIDIA says every major commercial-scale robotaxi program runs on its modular stack, spanning DGX training, simulation and validation, and in-vehicle computing. The robotaxi market is projected to reach $400 billion by 2035 with over 6 million commercial vehicles in operation.
BioIR accelerated protein-complex structure generation across 4,777 proteomes — about 31 million candidate complexes, with 1.81 million released as high-confidence predictions in the AlphaFold Database expansion. It keeps the PyTorch workflow, using optimized kernels and CUDA Graphs, with Ray running a model replica per GPU.
d-Matrix will connect its next-generation Raptor inference XPUs to NVIDIA's AI infrastructure via NVLink Fusion, MGX racks and Spectrum-X networking. NVIDIA cites 3x lower XPU-to-XPU latency than off-the-shelf Ethernet, 10x higher packet rates and 3 TB/s per XPU of all-to-all bandwidth on sixth-gen NVLink.
NVIDIA's blog details EPD disaggregation for multimodal inference, claiming up to 5x faster time-to-first-token and 7x faster end-to-end response. It separates vision encoding from prefill/decode, best for image-heavy prompts and quantized MoE models.
CUDA Toolkit 13.4 adds Windows on Arm support and early developer preview of the NVIDIA Rubin GPU architecture (compute capability 107). It also introduces MPS V3 with scriptable CLI, namespaces, and cgroup-integrated memory limits for shared GPU management.
At IBC 2026 in Amsterdam, NVIDIA announced a major expansion to NVIDIA AI for Media, adding SDKs, NIM microservices, and blueprints for broadcast, sports, and streaming. The Synthetic Video Detector (SVD) NIM now reaches 99.3% accuracy on text-to-video and 97.7% on image-to-video content.
NVIDIA announced CUDA Rust, enabling GPU kernels to be written in Rust and compiled natively to PTX. Two open-source projects from NVlabs: cuda-oxide for SIMT and cutile-rs for Tile. NVIDIA will grow CUDA Rust into 2027 and beyond.
Nvidia published an NVFP4-quantized build of Qwen3.8-27B on Hugging Face, drawing 65 likes and 10,488 downloads.
NVIDIA's blog details building a memory-driven Chief of Staff agent with NemoClaw, using a structured 'self model' knowledge layer. It shares five design lessons, including separating evidence, knowledge, and actions, and enforcing security boundaries with NVIDIA OpenShell.
NVIDIA's blog details how to deploy and optimize frontier reasoning and agentic AI models on Jetson edge devices, addressing size and performance challenges. It covers JetPack, containers, and inference pipelines for Jetson Orin and Thor.
NVIDIA Developer Blog explains how to propagate user identity across federated Kubernetes and AI platforms, covering scenarios like launching notebooks from a central portal. It details technical approaches for maintaining identity context in multi-cluster environments.
NVIDIA PAIR beta is available for Windows, macOS, and Linux, supporting GeForce RTX 20 Series and newer, RTX PRO, DGX Spark, and Apple M4+ silicon. It works with Ollama and LM Studio, routing inference across idle home systems without agent harness changes.
NVIDIA agreed to acquire Hugging Face for $12,930,300,000, per Jensen Huang's announcement. Hugging Face will stay an open platform supporting multi-cloud and multi-accelerator deployment, with NVIDIA compute not required to build on it.
NVIDIA's developer blog presents a step-by-step CUDA optimization walkthrough covering six incremental improvements, including CCCL API adoption, Compute Sanitizer, NVTX, CUB algorithms, pooled and pinned containers, and per-thread streams. Companion code and Google Colab option are provided.
NVIDIA's blog post details how speculative decoding accelerates LLM inference while maintaining accuracy, part of a series on AI model co-design. A companion video features Maor Ashkenazi explaining the technique.
At Fal.Con 2026, NVIDIA and CrowdStrike announced SafeMind, an agentic cybersecurity system built on NVIDIA Nemotron models. CrowdStrike reports its Blue Solano defensive model is 13% more accurate than the leading proprietary frontier model at 99% lower cost in internal evaluations.
NVIDIA's blog presents a practical framework for sizing GPU resources for AI inference workloads, focusing on use case, token patterns, latency targets, concurrency, cache hit rate, model choice, and deployment strategy. It emphasizes core-and-flex capacity planning and model optimization like quantization, pruning, and distillation to lower TCO.
NVIDIA and Anthropic integrated the BioNeMo Agent Toolkit into Claude Science, enabling agents to call BioNeMo NIM microservices for protein structure prediction. On internal benchmarks, BioNeMo skills raise task correctness from 60% to 100% and roughly double token efficiency.
NVIDIA Omniverse NuRec reconstructs real-world drives and renders new camera views for target vehicle configurations, enabling perception-stack adaptation without new datasets. It pairs reconstructed drives with target rigs, renders views, and refines frames with NVIDIA Harmonizer.
NVIDIA TensorRT Model Connect is an open collection of reference implementations that lets developers deploy supported open models from Hugging Face model ID to native C++ inference in two commands. It handles checkpoint mapping, engine construction, preprocessing, and post-processing.
AWS received its first NVIDIA Vera CPU server and Vera Rubin GPU, hand-delivered by NVIDIA's Ian Buck. Vera packs 88 custom Olympus cores, 1.2TB/s memory bandwidth, and up to 1.8x faster per-core performance on agentic AI workloads.
NVHBM integrates NVIDIA's memory controller into the HBM base die, delivering up to 30% greater memory bandwidth, 15% lower power consumption, and 25% more XPU compute die area vs. standard HBM4E. Amazon's Annapurna Labs will be first to work on NVHBM.
NVIDIA's tutorial applies an agent-driven COMPASS workflow to train cross-embodiment robot navigation policies, using Spot as reference and NVIDIA Omniverse NuRec for captured environments. It covers smoke testing, residual training, checkpoint evaluation, and runtime integration.
NVIDIA Dynamo's shadow engine recovery, now in preview, cuts LLM inference failover from 283 seconds to 7.3 seconds in a GLM-5.2 test. It keeps an idle initialized engine sharing weights via GPU Memory Service, so recovery happens off the serving path.
CUDA Python 1.0 ships with CUDA 13.3, offering stable APIs including cuda.core 1.0.0, cuda.compute 1.0.0, cuda.bindings 13.3.0, and nvmath-python 1.0. It gives Python developers full access to the CUDA platform without writing C++ extensions.
Alibaba released Qwen3.8-Flash-Next, a 176B-parameter multimodal MoE with 51B n-gram embeddings and 6B active per token, as a preview of the Qwen4 architecture. It beats Claude Opus 4.6 Max on 8 of 9 comparable benchmarks, including SWE-bench Pro and GPQA Diamond.
NVIDIA's Spectrum-X Ethernet Photonics is now in full production, delivering scale-out networking for AI factories with 4x fewer lasers and 5x lower power. The architecture co-designs switches and NICs to overcome traditional Ethernet's limitations for giga-scale AI training.
NVIDIA's NVLink Fusion integrates custom XPUs into its NVLink scale-up domain, targeting AI factory economics like tokens per second and cost per token. It aims to accelerate time to market and mitigate risk for semi-custom AI factories.
NVIDIA announced Groq 3 LPX, its interactive AI inference accelerator for the Vera Rubin platform, is in full production. In an Artificial Analysis benchmark running Gemma 4 31B, it delivered 3,431 output tokens per second on 100K context, 4x faster than the nearest alternative. Nebius is the first AI cloud to adopt it.
Measured with SemiAnalysis AgentX, Vera Rubin NVL72 achieves up to 30x higher throughput per megawatt than GB300 NVL72 on agentic workloads, and 35x lower token cost. Agentic requests consume 15x more tokens than chat, per OpenRouter data.
NVIDIA's Vera CPU is architected for agentic AI fleets, balancing per-core performance with burst absorption. Telemetry from 163,594 agentic sessions shows over 97% have unique trajectory profiles, making specialized CPU design points impractical.
NVIDIA introduced Scale-In, the fifth pillar of its AI networking, powered by BlueField-4 and DOCA over Spectrum-X Ethernet. It offloads security, storage, and data movement from host CPUs to dedicated DPUs for agentic AI factories.
Presents AdaptGrow, a GPU-accelerated SymNMF matrix factorization algorithm that turns rolling correlation and tail-dependence matrices into hard clusters, soft factor loadings, and structural-break signals at single-GPU and multi-node scale. A memory-efficient formulation cuts peak storage from ~20n items.
NVIDIA DSX MaxLPS is a suite of chip, thermal, system, and software technologies that maximizes AI factory throughput within a fixed power budget. It uses dynamic power allocation, software power optimization, and 45°C warm-water liquid cooling to cut cooling overhead and improve PUE.
NVIDIA's general-purpose coding agent AVO completed all 183 levels across all 25 public ARC-AGI-3 environments with no instructions, explicit rules, or stated goals. TechCrunch reports the same harness lifted Claude Opus 5 from 30% to 100% on the benchmark.
NVIDIA's AI safety teams outline the emerging agent stack—models, harnesses, meta-harnesses, secure runtimes like OpenShell, and inference infrastructure—and where security controls should live. They cite recent incidents where frontier agents from OpenAI, Anthropic, and the UK AI Security Institute operated beyond intended boundaries.
NVIDIA's hosted CUDA MCP Server gives AI coding agents one-line access to up-to-date CUDA documentation and code examples. The open-source Nsight Copilot Blueprint offers a self-hosted backend optimized for DGX Spark, with Nsight Compute integration providing guidance on issues like uncoalesced memory accesses.
NVIDIA's blog explains the shift from embedding-similarity to generative recommenders that predict the next item from user histories, addressing data volume, sparsity, and cold-start challenges. It highlights the recsys-examples and nv-embedding-cache tools for production-scale training and inference.
NVIDIA shows how a general-purpose coding agent (Codex with GPT-5.6 sol max mode) builds a real-time endoscopic tool segmentation app using Holoscan, HoloHub, and the Holoscan CLI. The workflow is agent-agnostic and iterative, with the engineer reviewing code and setting goals.
NVIDIA FLARE coordinates federated VLM training across data-local sites, using externalization, tensor streaming, and disk-backed aggregation for large updates. FedUMM, a William & Mary-NVIDIA project that federates lightweight adapters over a frozen multimodal backbone, won an Outstanding Student Paper Award at TheWebConf 2026.
NVIDIA released SkillEvaluator, an open-source tool that measures how skills affect agent performance via static checks and live task runs. First benchmark results cover 300+ verified skills across 30+ NVIDIA products, with plugins for Claude Code, Codex, and Cursor.
Cosmos 3 Edge is a 4B omni-model with a 2B Nemotron-based reasoner, pretrained on the same physical-world data as Cosmos 3 Nano and Super, and runs on-device on NVIDIA Jetson Thor. The tutorial covers post-training it to predict robot actions, serving policies on Jetson Thor, and evaluating in closed-loop simulation; the cosmos-framework repo and HuggingFace checkpoint are open.
NVIDIA's ALCHEMI Toolkit, introduced earlier this year, provides PyTorch-native building blocks for GPU-accelerated MLIP simulation workflows. This post demonstrates using AI coding agents with agent skills and reference files to generate simulation code, validated on NVIDIA H200 GPUs, with lessons from 45 generated pipelines.
Nemotron 3.5 Lightning pairs 30B LatentMoE parameters with 3B active, delivering up to 4x output speed and up to 1M context. The NVFP4 checkpoint compresses 66 GB to 22 GB; Perplexity Agent API pricing is $0.0115 input / $0.17 output per 1M tokens. Available open on Hugging Face and SageMaker JumpStart.
NVIDIA is partnering with SB Energy to secure LPS capacity at the PORTS-Pike Technology Campus in Portsmouth, Ohio, to host NVIDIA compute, with OpenAI as the tenant. The move applies NVIDIA's supply-chain discipline to secure infrastructure for frontier AI labs.
The coding-focused text-generation model is available on Hugging Face as nvidia/NVIDIA-Nemotron-Labs-Teacher-Competition-Coding, served via Transformers, vLLM, SGLang, or Docker with an OpenAI-compatible API.