NVIDIA AI Updates: September 24, 2026

1. NV-Reason-CT Brings Chain-of-Thought Radiology Reasoning to Full 3D CT Volumes

NVIDIA. NV-Reason-CT pairs a full 3D vision transformer encoder with a Qwen3.5-4B language model trained to produce reasoning traces that mirror how a radiologist works through anatomical regions. It reports 0.614 Macro-F1 and 0.871 Macro-AUROC on CT-RATE, beating published 3D contrastive and fused 2D/3D baselines, and NIH radiologists reviewed the structured reports for clinical plausibility. NVIDIA is releasing it as an open research foundation for post-training on specialized CT applications. Source

2. Transformer Engine Got Dropless MoE Training for JAX

NVIDIA. NVIDIA documented dropless mixture-of-experts training in JAX through Transformer Engine, targeting the routing path that DeepSeek, Qwen, and Mixtral-style architectures depend on. Dropless routing avoids discarding tokens when an expert’s capacity fills, which removes a quality-versus-throughput tradeoff that MoE trainers otherwise have to tune around. Source

3. SWE-Serve Measures Whether an Agent’s Patch Survives a Real Serving Load

NVIDIA. SWE-Serve evaluates changes to inference-serving software by checking whether a coding agent’s patch still works once a server loads a real model and handles live requests, not just whether unit tests pass. The gap it exposes is specific to systems code: a patch can be locally correct and still break under concurrency, memory pressure, or model loading. Source

4. NVIDIA Argued Health Checks Do Not Prove a GPU Cluster Can Run a Job

NVIDIA. A cluster where every GPU, network link, and pod reports healthy can still fail a 512-GPU training run, and NVIDIA published a validation methodology for catching that before workloads land. The distinction is between component-level health and end-to-end collective performance, which only shows up under an actual distributed run. Source

5. NodeWright Manages the Node Itself, Not What Runs on It

NVIDIA. NodeWright handles kernel settings, system packages, storage layouts, and security agents across Kubernetes node fleets, filling the gap Kubernetes leaves by managing workloads rather than the machines under them. For GPU fleets where driver and kernel drift silently degrades throughput, that layer has usually been a pile of bespoke Ansible. Source