Breaking

NVIDIA Shows Cosmos 3 Nano Fine-Tuning Lifts Traffic-Scene Accuracy From 54% to 93% in Under a Day

July 14, 2026 at 13:00 EDT

Foundation Models
Robotics
Open Source

NVIDIA has demonstrated that its open Cosmos 3 Nano world foundation model can be adapted to a specialized traffic-recognition task and lift validation accuracy from roughly 54% to 93% in less than a day, using a combination of lightweight fine-tuning and automated model tuning.

Physical AI · Cosmos 3 Nano

From 54% to 93% Accuracy in Under a Day

NVIDIA adapted its open Cosmos 3 Nano world model to a traffic-signal recognition task — using lightweight LoRA fine-tuning plus automated TAO tuning to nearly double its validation accuracy in less than 24 hours.

+39pts

Accuracy gain, zero-shot → fully tuned

<1 day

Full adaptation cycle, start to finish

~16B

Parameters, workstation-class model

WTS Validation Accuracy Across Three Stages

Each stage layered on the previous one

54.41%

Zero-shot

→

87.14%

+ LoRA
post-training

→

93.35%

+ TAO
AutoML

1 · LoRA

Updates only a small subset of parameters — far cheaper than full retraining.

→

2 · TAO AutoML

Automates hyperparameter search to squeeze out extra accuracy — no manual tuning.

WHY IT MATTERS

A general physical-AI model specialized for a concrete perception task in a sub-day cycle — evidence NVIDIA's stack can be productized fast.

OPEN QUESTIONS

The WTS validation set is not detailed publicly, making the numbers hard to compare — and generalization to diverse real-world driving remains untested.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.