NVIDIA-NeMo

cuML and cuVS 25.06 distribute UMAP's all-neighbors kNN construction across GPUs by partitioning into balanced clusters, computing local graphs, then merging — no all-to-all communication. Result: 870GB of vectors in 8 minutes on eight H100s, up to 74x faster than projected CPU runs.

NVIDIA ships OpenShell 0.1.0, an open-source runtime that restricts which systems and data an agent can reach via sandboxing, credential isolation, and a formal policy prover, without modifying agent code. Cadence, Slack, and Gecko Robotics are early adopters.

NVIDIA and Groq detail how cycle-exact deterministic scheduling across 256 LPU chips enables power-smoothing tricks that cut voltage guardbands, claiming up to 35x more inference throughput per megawatt than GB200 NVL72.

NVIDIA details a Transformer Engine recipe for training mixture-of-experts biological foundation models, combining GroupedLinear batched expert GEMMs, MXFP8 quantization, and kernel fusion to hit 2.21x the throughput of a Hugging Face baseline on 8xB200 GPUs.

NVIDIA engineers give rules for tuning speculative decoding: set draft length to 128/G minus 1 when attention dominates, align longer drafts to 128-element tile boundaries, keep GEMMs compute bound without adding KV cache pressure, and extend drafts only while acceptance gains beat draft cost. Compa

Adapting vision reasoning models to production video tasks usually costs days in data formatting, container setup and training scripts. This uses agent skills to compress that into a single day of post-training.

Coding agents can now inspect repositories, set up runtimes, resolve build failures, launch experiments, monitor them and summarise results, which makes them practical operators for long-running machine learning workflows.

A sizing method for inference fleets: classify the workload by token pattern, then derive memory and compute from daily actives, concurrency, input and output lengths, cache hit rate and latency targets. Reserved capacity carries baseline traffic while spot GPUs absorb bursts.

Every replica cold start, rolling update and RL weight refit re-pays the cost of moving a checkpoint. ModelExpress asks where a compatible copy already lives and transfers it GPU-to-GPU over RDMA, falling back to streamed object storage, cutting a measured startup from eight minutes to under two.

NVIDIA breaks down why MoE models like Nemotron 3.5 Lightning decouple memory from compute cost, when that beats a dense model like Gemma 4 31B, and where quantization and fine-tuning need extra care in MoE routers.

NVIDIA details RTX Spark: GB10 with up to 128GB unified memory, first PyTorch-CUDA build for Windows on Arm, a 2-bit 74B MoE model for local agents, and vendor-claimed gains over Intel Panther Lake in agent and fine-tuning workloads.

NVIDIA's VSS Blueprint 3.3 adds a Build Vision Agent skill for prompt-based deployment and Adaptive Efficient Video Sampling, reporting 17% lower alert latency and 46% more concurrent real-time VLM streams on one GPU.

NVIDIA's Groq 3 LPX, now in production on Vera Rubin NVL72, posts 3,431 output tokens/sec at 100K context on Gemma 4 31B, using compiler-scheduled execution and preplanned chip-to-chip networking to keep tensor parallelism efficient at small batch sizes.

NVIDIA and Nscale benchmark DSX MaxLPS, a dynamic power-sharing scheme that packed 40% more GPUs into a fixed power budget and raised throughput per watt from 4.10 to 6.12 tokens/s/W, at the cost of a 17% rise in P99 latency.

A working end-to-end post-training loop on a small open model: establish a baseline, run reinforcement learning with verifiable rewards against a hosted environment, get back a downloadable LoRA adapter, then re-evaluate. Setup is minutes, and the same recipe scales to the larger models in the famil

NVIDIA details official eight-node DGX Spark clustering with 1 TB of unified memory, a Spark-tuned collective backend, decode and training scaling numbers, switch versus ring topologies, and a router that splits traffic across model pools.

NVIDIA describes cuDF skills for Spark that guide coding agents through testing, converting, benchmarking and optimizing UDFs. Rewrites use SQL expressions, RapidsUDF or CUDA kernels, and one production workload dropped from 274.55 to 12.78 operator hours.

An NVIDIA technical blog post examining why a single shared Kubernetes cluster becomes hard to manage as more teams onboard, citing conflicting CRD versions, overlapping RBAC policies, and the lack of a clean mechanism to partition GPU capacity into team-level budgets. It discusses multi-tenancy approaches for isolating teams on shared GPU infrastructure without provisioning a fully separate cluster per team.

NVIDIA breaks down how Qwen3.8-Flash-Next attacks long-context inference: Gated DeltaNet compresses history while Qwen Sparse Attention aggregates tokens into micro-blocks for block-level retrieval. Measured against full attention, prefill runs 7.6x faster and decode 4.9x, at 16K+ tokens/sec per GPU

NVIDIA and CrowdStrike ran offensive and defensive agents as a closed loop in an isolated lab, with Nemotron 3 Ultra orchestrating defense and a post-trained Nemotron 3 Super writing detections against Falcon telemetry. Backtests reached a 41.9% mean detection rate, 2.5x the default harness.

NVIDIA reconstructs real driving scenes with 3D Gaussian splatting, re-renders them for whatever camera rig a new vehicle uses, refines the frames with Harmonizer, and trains perception models on the output. The scene dataset and an agent-skills repo for the workflow are public.