System Card
← leaderboard
NVIDIA
model systemcard
Nemotron 3 Ultra
#23
of 27 · general
550B A55B; the Nano/Super rungs of the same ladder stay off the board
550B
params · MoE
20T
training tokens
OpenMDW License v1.1
license
2026-06-04
released
2025-09-30
knowledge cutoff
paper ↗
weights ↗
repo ↗
Standing
dot = this model · shaded = the field's middle 60% · track = full field
General
9.4
#23
/27
Math
12.0
#24
/27
Coding
6.2
#26
/27
Health
0.0
#26
/26
Engineering
0.0
#17
/17
Analyst
0.0
#11
/11
Legal
55.0
#9
/26
Instruct
29.7
#19
/27
Agents
15.9
#23
/27
Professional
13.2
#23
/27
Hallucination
26.9
#14
/27
Speed
56.7
#11
/22
Value
24.1
#23
/27
Open
36.8
#5
/11
Measured
≈ imputed · amber dot = lab-self-reported among the sources
GPQA Diamond
86.7%
#22
HLE
32.9%
#22
MMLU-Pro
86.3%
#17
SWE-bench V
69.8%
#26
LiveCodeBench
86.0%
#11
SciCode
39.9%
#24
CritPt
3.1%
#23
Terminal-Bench
52.4%
#23
APEX Agents
11.5%
#17
GDPval
33.1%
#19
Arena
1426
#22
AA Index
38.3
#21
Vals Index
44.0
#21
Epoch Capabilities Index
142.8
#23
Operations
$0.60
/
$2.75
$ per 1M tokens, in / out
$1.14
blended
144
tokens / sec
1426
arena elo · 10,583 votes
Neighbors
nearest capability fingerprints · deltas describe the neighbor
Tencent
HY3
→ stronger professional · cheaper
MiniMax
MiniMax M2.7
→ less factual · looser instructions
DeepSeek V4 Flash
→ cheaper
Xiaomi
MiMo V2.5 Pro
→ stronger coding · cheaper
TM
Inkling
→ pricier