AI news for builders and product teamsUpdated Nov 17, 2026, 13:01 UTC
NVIDIA Developer Blog
First-party releases and research from NVIDIA Developer Blog. Headlines and excerpts link to the original articles.
Latest stories
Newest first
NVIDIA and Princeton's Dao AI Lab have released ReplaySSM, a decoding technique that caches recent state-space model inputs instead of full recurrent state, cutting SSM state memory traffic roughly in half and speeding up hybrid-model decoding. It is already integrated into TensorRT-LLM, vLLM, SGLang and FlashInfer and runs in production in GLM-5.3-Flash and Kimi-K3.

NVIDIA detailed a five-step workflow for turning CAD files into simulation-ready OpenUSD assets for robotics, demonstrated on an ABB YuMi robot with assistance from a frontier AI model. The validated asset completed eight gripper pick-and-place cycles in Isaac Sim.

NVIDIA's KGMON team placed second in the KDD Cup 2026 Data Agents competition with a system built on a small, fixed LLM and a deliberately minimal harness. The team's approach centered on constraining the agent's action space and making every execution attempt inspectable.

NVIDIA's Seattle Robotics Lab and Isaac team built robots to assemble GB300 tester trays, reaching over 95% success on busbar assembly with a 160-second cycle time, though short of the 124-second target.

NVIDIA published a developer blog describing how its DSX Air node-based digital twin and Brev GPU compute can be combined with AI agents to validate AI factory infrastructure and software changes before production. The post also cites the NVIDIA AI Blueprint for Video Search and Summarization as a demonstration of the pattern.

NVIDIA introduced the mPDLP solver in cuOpt, distributing large linear programming problems across NVLink-connected GPUs with up to 6x lower peak memory per GPU versus single-GPU PDLP and LP problems capped at 2.1B nonzeros. Kinaxis and PSR reported speedups on large models.

NVIDIA announced cuPhoton, an open-source CUDA-X toolkit that keeps scientific image data on the GPU from sensor read through classification. NVIDIA reports image loading speedups up to 14,900x and signal processing up to 14,550x versus an x86 CPU baseline on representative multi-GPU workloads.

NVIDIA reports that a workflow proposed in Megatron-LM PR #7262 uses ordered per-rank tensor fingerprints to locate nondeterminism in large-scale pretraining, and that kernel and recipe optimizations cut the determinism tax to roughly 2% at 2,432 GPUs while holding bitwise determinism over 800 steps.

NVIDIA's DOCA GPUNetIO provides a unified GPU-initiated networking foundation for CUDA kernels to drive Ethernet, RDMA, Verbs, and DMA operations without the CPU on the critical path. It ships as both a full DOCA SDK superset and a lighter open-source Verbs-focused implementation, and is now used by NCCL, NVSHMEM, UCX/NIXL, and Holoscan Sensor Bridge.

NVIDIA released AICR v1.0, its AI Cluster Runtime, which provides version-locked, validated recipes for configuring GPU-accelerated Kubernetes clusters. The release establishes a stable compatibility contract across the CLI, REST API, Go SDK, bundle layout, and artifact schemas, with signed validation evidence.

NVIDIA's green contexts, available in the Driver API since CUDA 12.4 and now in the Runtime API starting with CUDA 13.1, let applications select subsets of GPU execution resources and target work to them explicitly. NVIDIA describes SM partitioning, workqueue provisioning, and an example comparing critical kernel latency across three execution modes on a Blackwell GPU.

NVIDIA released DOCA AI agent skills on GitHub that provide verified API signatures, hardware capability requirements and build constraints for BlueField DPU development. In a 65-prompt evaluation, agents without skills satisfied 19% of checklist items versus 100% with skills.

NVIDIA's Do Inference Now (DIN) Deploy is an open-source collection of C++ samples combining ONNX Runtime with the TensorRT RTX execution provider for local AI inference on Windows and Linux. It covers speech recognition, image segmentation and image generation, with DGX Spark performance measurements included.

NVIDIA detailed a NeMo framework workflow for fine-tuning Nemotron 3.5 ASR on Saudi Najdi and Hijazi dialects using 133.7 hours of speech, reporting word error rate dropping from 55.05% to 29.96% on the target test split. NVIDIA also released Nemotron 3 Diarization for speaker-attributed transcription of up to eight speakers.

NVIDIA detailed an end-to-end HSTU generative recommender inference workflow in its recsys-examples repository, served through NVIDIA Dynamo-Triton with PyTorch AOTI and FlexKV-backed KV caching. On an RTX PRO 6000 Blackwell Workstation Edition GPU, the eight-layer HSTU model reached up to 5.93x lower latency at batch size 8 with a 100% GPU KV-cache hit rate versus the same AOTI configuration without caching.

NVIDIA announced general availability of its cuObject client and server libraries and expanded the xio-sig consortium to include cuObject alongside cuFile, with Google Cloud evaluating participation and Microsoft planning to join the board. A new SCADA Server SDK lets storage providers build servers for GPU-initiated requests, and IBM demonstrated a prototype integrating SCADA with IBM Storage Scale.

NVIDIA published a tutorial on using NeMo Relay execution traces to inspect Hermes Agent behavior, including two runnable examples and a ToolPerf case study comparing baseline and fixed harness revisions across 108 runs.

NVIDIA published lessons learned from building TensorRT Model Connect, an open source collection of AI model reference implementations in C++ on TensorRT. As of the July 29, 2026 release, it covered 128 model families tested on NVIDIA GB300 and is in public preview.

NVIDIA released VSS Blueprint 3.3, which adds a Build Vision Agent skill that composes multi-workflow visual AI deployments from a single prompt and Adaptive Efficient Video Sampling (EVS) to reduce runtime VLM processing. NVIDIA reports a bottling-line demo deployed in under 30 minutes and measured latency and concurrency improvements on RTX PRO 6000 Blackwell hardware.

NVIDIA introduced its Open Agent Safety Platform, combining the open-source OpenShell secure runtime with NVIDIA Sentry on BlueField-4 DPUs to monitor and enforce policies for AI agents. It is built around five stated principles and is optimized for NVIDIA Vera CPU and BlueField DPU systems.