Breaking

NVIDIA Nemotron 3 Ultra Lands in Open-Weight Top Tier on 3D Design Arena

June 25, 2026 at 17:35 EDT

Foundation Models
Open Source
AI Agents

NVIDIA's open-weight model Nemotron 3 Ultra, released in June 2026, has ranked among the top nine open-weight models in the 3D category of the design-generation benchmark Design Arena. The result shows the model delivering strong performance while keeping an average input cost around the $0.50 class, among the most cost-efficient in its tier.

June 25, 2026 · NVIDIA · 3D Design Arena

Nemotron 3 Ultra: a top-9 open-weight model that wins on cost

NVIDIA's open-weight frontier model ranks among the top nine labs on a human-voted 3D design benchmark — and matches closed rivals at a fraction of the token cost, drawing on a 550B-parameter Mixture-of-Experts with a Mamba-Attention hybrid.

$0.50

Low average input cost on the design benchmark

Top 9

Open-weight labs on 3D Design Arena

550B

Total params (55B active, MoE)

Token context window

Same quality, one-tenth the cost

HTML5 canvas + physics simulation: Nemotron 3 Ultra matched GPT-5.5 in quality while spending far fewer tokens.

$0.051

Nemotron 3 Ultra

$0.57

GPT-5.5

≈11× cheaper per task for comparable output

Inference throughput advantage

At 8k input / 64k output, throughput relative to rival open models.

5.9×

vs GLM-5.1

4.8×

vs Kimi-K2.6

1.6×

vs Qwen-3.5

Closing in on closed models in law

Legal Agent Benchmark all-pass rate after fine-tuning — landing between two leading closed models, at one-eighth to one-fiftieth their cost.

4.2%

Sonnet 4.6

5.8%

Nemotron 3 Ultra

6.6%

Opus 4.6

Reliability (rubric pass rate) climbed from the low 70s to near 95% — up from a 0% all-pass baseline.

Continue reading

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