Inference
EN

Nemotron 3.5 Lightning 30B A3B NVFP4 vs Nemotron 3 Ultra

Change models
Nemotron 3.5 Lightning 30B A3B NVFP4NVIDIA
Nemotron 3 UltraNVIDIA
IndexAll categories, out of 100.
37.0 ±6.8
50.8 ±5.3 best
GapIndex points below the top.
−13.8
Top
CoverageIndex weight that has results.
75%
80% best
SpeedOutput tokens per second.
—
—
EloLMArena rating and rank.
—
—
ReleasedMonth of release.
2026-08
2026-06
Input / 1MUS dollars per 1M input tokens.
—
Free
Output / 1MUS dollars per 1M output tokens.
—
Free
1K chats2K tokens in, 500 out, each.
—
Free
1K documents20K tokens in, 1K out, each.
—
Free
1K agent tasks200K tokens in, 20K out, each.
—
Free
ContextMaximum tokens in one request.
1M
1M
AgenticMulti-step tasks with tools.
30.2
48.9 best
CodingCode writing and repair.
38.7
54.3 best
ReasoningLogic problems and puzzles.
38.9
50.0 best
MultimodalTasks with images and text.
—
—
KnowledgeFacts and expert knowledge.
39.5
50.0 best
MultilingualTasks in many languages.
—
52.4
InstructionTasks with strict prompt rules.
57.1
65.2 best
MathMath problems.
—
—

Benchmarks

19 shared
Only Nemotron 3 Ultra: 13
Nemotron 3.5 Lightning 30B A3B NVFP4
Nemotron 3 Ultra
Artificial Analysis Humanity's Last ExamKnowledge
10.6%143 / 222
28.4% best89 / 222
Artificial Analysis Long Context ReasoningReasoning
49.2%159 / 217
67.0% best130 / 217
Critical Physics TasksReasoning
0.0%150 / 214
3.1% best95 / 214
Artificial Analysis GPQA DiamondKnowledge
74.3%122 / 199
86.7% best62 / 199
Artificial Analysis Omniscience AccuracyKnowledge
14.7%153 / 164
21.6% best123 / 164
Graduate-Level Google-Proof Q&AKnowledge
75.6%78 / 141
87.0% best44 / 141
GDPval-AA normalizedAgentic
6.9%93 / 133
33.1% best55 / 133
Artificial Analysis SciCodeCoding
32.1%102 / 114
40.3% best86 / 114
Software Engineering Benchmark VerifiedCoding
52.8%94 / 114
71.9% best63 / 114
Terminal-Bench 2.1 · BenchLMAgentic
23.5%82 / 86
56.4% best68 / 86
GPQA Diamond · BenchLMKnowledge
75.6%68 / 84
87.0% best49 / 84
Massive Multitask Language Understanding ProfessionalKnowledge
81.6%36 / 75
86.8% best9 / 75
BrowseCompAgentic
36.8%62 / 63
44.4% best61 / 63
Instruction Following BenchmarkInstruction
72.9%37 / 57
81.7% best5 / 57
Humanity's Last Exam without toolsKnowledge
10.5%51 / 53
26.7% best44 / 53
Scientific Code BenchmarkCoding
31.4%30 / 35
44.6% best15 / 35
τ³-Bench Tool-Agent-User EvaluationAgentic
9.5%30 / 32
70.9% best4 / 32
PinchBenchAgentic
83.4%6 / 10
90.0% best3 / 10
Artificial Analysis Openness IndexKnowledge
83.3%
83.3%