Breaking

NVIDIA's Nemotron 3 Ultra Ranks Among Top 9 Open-Weight Labs in 3D Design Arena

June 25, 2026 at 17:33 EDT

Foundation Models
Open Source
AI Agents

NVIDIA's open-weight model Nemotron 3 Ultra has been placed among the top nine open-weight labs in the 3D category of Design Arena, a benchmark that compares AI-generated design output. It earned the ranking by delivering strong performance while keeping an average input cost of just $0.50.

June 4, 2026 · NVIDIA

Nemotron 3 Ultra: NVIDIA's Open-Weight Flagship for Long-Running Agents

The 550B/55B-active MoE model ranks 9th among open-weight models on 3D Design Arena — and lands as the top-ranked fully-open U.S. entry, with weights, data, and recipes released.

1213

Elo on 3D Design Arena (human-voted)

550B/55B

Total / active params (MoE)

Token context length

~70

SWE-Bench Verified score

Agentic efficiency gains

Up to 5× faster inference vs. baseline on long-running agent workloads

1×

baseline

5×

Nemotron 3 Ultra

Plus up to 30% lower inference cost — inference speed stays linear across hundreds of agent steps.

Pricing per 1M tokens

$0.50

Input

$2.20

Output

$0.10

Cached input

Hybrid Transformer + Mamba-2 architecture and Native NVFP4 for Blackwell deliver low-memory, high-speed inference — positioning it among the most cost-efficient models in its class.

Strengths

Fully open: weights, data & training recipes
Top-ranked U.S. open-weight design entry
Linear speed across long agent loops
Cuts task-completion time in coding agents

Trade-offs

High memory needs for local inference
Fine-grained design tasks may need fine-tuning
Chinese models (Qwen, GLM, Kimi) still lead the open field

Continue reading

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