NVIDIA AI Updates: August 13, 2026

1. NVIDIA Details Serving Qwen3.8-2.4T-A95B on GB300 NVL72

NVIDIA. NVIDIA published a deployment path for serving Alibaba’s Qwen3.8-2.4T-A95B, an open-weight mixture-of-experts model with 2.4 trillion total parameters and 95 billion activated per token, on GB300 NVL72 systems. The model uses hybrid attention that alternates full-attention and linear-attention layers to hold contexts up to one million tokens while bounding KV cache growth, and NVIDIA reports over 4,000 tokens per second per GPU and more than 350 tokens per second per user at FP8 precision without additional tuning. NVIDIA offers serving through SGLang, vLLM, and NVIDIA Dynamo, a model-free NIM container for streamlined deployment, and fine-tuning via NeMo AutoModel, with configurable per-request reasoning controls. Source

2. NVIDIA Releases JetPack 7.2.1 With Agentic Video Skills and T3000 Emulation

NVIDIA. NVIDIA released JetPack 7.2.1 for its Jetson edge platform, adding agentic video skills that translate high-level goals into verified codec configurations by automating device discovery, capability checks, recipe generation, and benchmarking. The update brings PyNvVideoCodec 2.2 to Jetson for the first time, providing hardware-accelerated, GPU-resident video encoding and decoding through DLPack, plus threaded decoding that decouples decode latency from inference. It also lets developers emulate the newly announced Jetson T3000, rated at 865 FP4 TFLOPS, on a Jetson T5000 module so they can build on the Thor AGX Developer Kit before deploying to the more power-efficient part. Source

3. NVIDIA Spotlights Local AI Tooling for Open Models and Agents

NVIDIA. NVIDIA published a roundup of its work with the local AI community to run open models and agents on personal hardware, highlighting new developer tooling alongside recently released models. It featured Unsloth Desktop, described as the first desktop application that both trains and runs AI models locally, and NVIDIA Sync clustering that lets multiple DGX Spark systems operate together. The post reiterated support for open models across consumer RTX GPUs, DGX Spark, and Jetson devices, along with the NeMo Switchyard routing library for directing workflow steps to the most suitable model. Source