Breaking

NVIDIA's Nemotron 3 Ultra Lands 9th on 3D Design Arena

June 25, 2026 at 17:38 EDT

Foundation Models
AI Agents
Open Source

NVIDIA's open-weight model Nemotron 3 Ultra has placed 9th among open-weight labs on "3D Design Arena," a benchmark scored by human preference voting. With an average input cost of about $0.50, the model delivered strong performance at low cost, ranking just behind the leading group's Alibaba Qwen.

December 2025 · NVIDIA

Nemotron 3 Ultra cracks the open-weight top nine — at $0.50

NVIDIA's flagship open model lands among the leaders on the 3D Design Arena, trailing only Alibaba Qwen — and it does so at one of the lowest input costs in its class.

Top 9

open-weight labs on 3D Design Arena

$0.50

average input cost — low for its class

550B

total params · ~55B active (MoE)

4M+

creators behind the benchmark

Efficiency gains on agentic tasks

Reported versus other open frontier models — the size of each column is the figure.

5×

faster inference

1×

baseline (other open models)

−30%

lower cost on agents

Fully open by design

Open weights

BF16 & NVFP4 versions published

Open data

Training data released

Open recipes

Run via NIM, vLLM, OpenRouter, DeepInfra

Well received

Strong efficiency-to-capability balance — the earlier Super ran on 2× H100 at 484 tok/s with results rivaling much larger models. Easy local and cloud deployment via Hugging Face and vLLM.

Caveats

The large MoE carries high VRAM requirements, and detailed independent reviews remain limited this soon after release.

Continue reading

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