Skip to main content
Language modelActive

Nemotron 3 Ultra

NVIDIA

Released
June 4, 2026
Data date
September 8, 2026
Intelligence
49.8 / 100
Overall score
Speed
162.9 tokens/s
Output throughput
Price
$0.6 · $2.4
Input · Output / 1M tokens
Input
Text
Output
Text

Nemotron 3 Ultra combines 550 billion total parameters with 55 billion active parameters in NVIDIA's LatentMoE architecture. Mamba2, MoE, attention, and multi-token prediction form the technical basis. The checkpoint supports ten languages and returns text only.

The 1M-token window covers long software and document tasks. NVIDIA's technical report describes the efficiency approach without making a blanket quality promise. For memory and parallelism, the full 550-billion-parameter footprint still matters.

Position
Rank 38 of 86
Index score
49.8 / 100
1.Claude Fable 5.193.5 / 100
2.Claude Fable 593.2 / 100
3.Claude Opus 591.3 / 100
4.Claude Sonnet 5.587.9 / 100
5.Claude Opus 5.585.1 / 100
6.Kimi K385.1 / 100
7.Gemini 3.7 Flash83.1 / 100
8.GPT-5.6 Sol81.9 / 100
9.Claude Opus 4.878.8 / 100
10.Gemini 3.8 Flash77.4 / 100
11.Qwen 3.8 Max 090276.5 / 100
12.GPT-5.575.8 / 100
13.Claude Opus 4.772.6 / 100
14.Gemini 3.6 Flash70.4 / 100
15.Grok 4.669.7 / 100
16.Qwen 3.7 Max68.7 / 100
17.GPT-5.268.3 / 100
18.Gemini 3.1 Pro Preview67.5 / 100
19.GPT-5.6 Terra67.5 / 100
20.Gemini 3.5 Flash67.4 / 100
21.Gemini 3 Pro66.3 / 100
22.GLM-5.365.3 / 100
23.GPT-5.6 Luna64.4 / 100
24.Grok 4.763.7 / 100
25.DeepSeek-V4.1-Flash62.8 / 100
26.Gemini 3 Flash Preview62.7 / 100
27.GPT-5.462.7 / 100
28.GLM-5.3-Flash62.4 / 100
29.Claude Sonnet 561.6 / 100
30.Kimi K2.659.2 / 100
31.Grok 4.558.6 / 100
32.MiniMax M358.3 / 100
33.GPT-557.6 / 100
34.Claude Sonnet 4.655.9 / 100
35.Qwen 3.8 27B53.7 / 100
36.GLM-5.252.6 / 100
37.GPT-6 Luna50.7 / 100
38.Nemotron 3 Ultra49.8 / 100
39.GPT-6 Sol49.6 / 100
40.DeepSeek-V4-Pro49.6 / 100
41.Inkling Small48.5 / 100
42.Kimi K2.548.2 / 100
43.Inkling47.8 / 100
44.o346.9 / 100
45.GPT-5 mini46.1 / 100
46.GLM-546 / 100
47.GLM-5.144.2 / 100
48.Claude Opus 4.543.7 / 100
49.GPT-5.4 mini42.3 / 100
50.MiMo-V2.5-Pro42.2 / 100
51.Grok 4.20 Reasoning39.3 / 100
52.Grok 338.7 / 100
53.Gemini 3.1 Flash-Lite38.5 / 100
54.Gemini 3.5 Flash-Lite38.2 / 100
55.Grok 4.336.8 / 100
56.MiniMax M2.135.5 / 100
57.MiniMax M2.735 / 100
58.MiMo-V2.533.6 / 100
59.GLM-4.733.4 / 100
60.o4-mini33.3 / 100
61.GPT-4.132.6 / 100
62.Ling 3.0 Flash32.2 / 100
63.o131.4 / 100
64.GPT OSS 120B30.6 / 100
65.Kimi K2 Thinking28.8 / 100
66.MiniMax M2.528.2 / 100
67.DeepSeek-R127.3 / 100
68.GLM-4.625.8 / 100
69.GLM-4.525.6 / 100
70.Kimi K2 071124.4 / 100
71.GPT-5.4 nano24 / 100
72.DeepSeek-V3.222.4 / 100
73.Mistral Large 319.8 / 100
74.GPT-4.1 mini19.7 / 100
75.o3-mini19.7 / 100
76.DeepSeek-V3 032418.8 / 100
77.DeepSeek-V317.8 / 100
78.GPT OSS 20B15.1 / 100
79.GPT-5 nano12.4 / 100
80.Command A10.5 / 100
81.Mistral Medium 3.59.5 / 100
82.Llama 4 Scout6.1 / 100
83.Mistral Small 3.14.7 / 100
84.GPT-4.1 nano4 / 100
85.Command R+2.5 / 100
86.GPT-4o mini2.3 / 100
Methodology and source

The index averages rank percentiles from 10 documented comparison cohorts.