Nvidia

At IFA 2026, Nvidia, Microsoft, and partners announced a set of tools to make local AI inference faster and easier to deploy on Nvidia hardware, headlined by NVIDIA RTX Spark compact Windows PCs arriving in October from Lenovo and Acer. The company also introduced NVIDIA PAIR, a Personal AI Router that distributes AI i The blog also notes simplified local AI support in Hermes Agent, OpenClaw, and Perplexity Portable Computer, and highlights a wave of August local-AI model releases runnable on RTX PCs, DGX Spark, DGX Station, and Jetson, including Nvidia's own Nemotron 3.5 Lightning (30B), Z.ai's GLM-5.3-Flash, Qwen3.8-Flash-Next, Qwe

NVIDIA and Nebius Group N.V. (NASDAQ: NBIS) today announced a strategic partnership to develop and deploy the next generation of hyperscale cloud for the AI market, from AI natives to enterprises.

NVIDIA's official Newsroom index (July 14–17, 2026) surfaces multiple provider-relevant announcements. On July 17, 2026, NVIDIA detailed how its Vera Rubin platform delivers the lowest cost per token for post-training workloads, framing "intelligence per dollar" as a key metric for agentic AI and highlighting extreme c The same newsroom window adds further provider-facing items: on July 15, 2026, NVIDIA introduced new Jetson Thor computers aimed at mainstream robotics and edge AI, positioning compact, power-efficient AI supercomputers for running foundation models at the edge, and separately announced that Japanese enterprises and st

NVIDIA published vLLM Release 26.06, a container image available on NGC that bundles the upstream vLLM serving engine with NVIDIA-tuned libraries for GPU-accelerated LLM inference. The release includes vLLM 0.22.1, flashinfer 0.6.12, transformers 5.6.0, flash-attention 2.7.4.post1, xgrammar 0.2.0, and Torch 2.13, built The release notes also document known issues relevant to developers running on shared-memory platforms: default aggressive GPU memory allocation can cause out-of-memory errors on DGX Spark and Jetson, Nemotron Nano V3 and Nemotron Super V3 NVFP4 models on Spark require limiting `--max-num-seqs` to 4, and gpt-oss-120b o

Nvidia's NIM for Large Language Models and Vision Language Models release notes for version 2.0.12 formally expand the product to include VLM support alongside LLM inference, refreshing model-specific and model-free containers to 2.0.12 and updating the inference backend to vLLM 0.27.1 for both GPU and CPU images. The Bug fixes in 2.0.12 include a hybrid Mamba prefix-cache fix that preserves a prefix-cache hit when speculative decoding resumes, restoring time-to-first-token on profiles like nemotron-3-ultra-550b-a55b, and a Helm chart fix that applies the configured service account to multi-node deployment pods. These are first-part