NVIDIA AI Updates: May 30, 2026

1. DynoSim Lets Teams Screen Inference Deployments 1,500x Faster Than Real Time

NVIDIA. NVIDIA released DynoSim, a Rust discrete-event simulator that models the full Dynamo LLM serving stack on a single virtual timeline, covering request routing, scheduler decisions, forward passes, and KV cache management. It composes workload replay, single-engine simulations, the Router, the Planner, and KV management as separate actors on one event queue, and pulls measured engine timing from AI Configurator plus backend-specific scheduler logic for vLLM and SGLang. The tool replayed the 23,608-request Mooncake trace in 2.41 seconds on an M4 MacBook Air, roughly 1,500x faster than real time, so teams can sweep tensor-parallel shapes, worker counts, routing policies, and autoscaling settings before spending GPU hours on validation. Source

2. StepFun’s Step 3.7 Flash Lands on NVIDIA NIM With Day-0 Inference Support

NVIDIA. NVIDIA detailed enterprise deployment of StepFun’s Step 3.7 Flash, a multimodal vision-language model with a 198B-parameter Mixture-of-Experts design that activates roughly 11B parameters per forward pass across 288 experts. The model adds native image and video input, three configurable reasoning levels, and a 256k-token context window, packaged through NVIDIA NIM with an OpenAI-compatible API and support for SGLang, TensorRT-LLM, and vLLM. NVIDIA reports about 600 tokens per second on Hopper GPUs and offers Day-0 fine-tuning via NeMo, positioning the release for financial analysis, coding agents, and high-throughput multimodal workloads. Source

3. MCG Toolkit Automates Model Card Generation for AI Governance

NVIDIA. NVIDIA introduced the Model Card Generator (MCG) Toolkit, a containerized pipeline that reads source code and associated files to produce Model Card++ documentation, an overview plus Bias, Explainability, Privacy, and Safety and Security subcards, in under a minute. The three-stage flow ingests content from GitHub, GitLab, Hugging Face, URLs, or uploads, uses Nemotron RAG for embeddings and reranking with GPT-OSS-120B for generation, then renders structured JSON into Markdown. NVIDIA reports 91% completion and 76% accuracy on standardized tests, framing the tool as a way to keep documentation current with regulatory regimes like California AB-2013 and the EU AI Act. Source