Atlas

Models

← All models

Nemotron 3.5 Lightning 30B A3B BF16

NVIDIA · 2026-08-11 · 32.9B parameters

NVIDIA's BF16 reference release of Nemotron 3.5 Lightning, a hybrid Mamba-2, attention and mixture-of-experts language model with optional thinking. The official card documents a 1,048,576-token maximum context and an August 11 release; Fireworks explicitly links its nemotron-lightning-3p5-30b-a3b endpoint to this checkpoint. NVIDIA reports approximately 3B active parameters and NVFP4 pretraining; BF16 describes the distributed weights and is not silently assigned as the training precision.

Benchmark scores

BenchmarkScore
Capability133.1
Code Migration3.0
Code Migration - CLI3.9
Code Migration - COBOL0.0
Code Migration — Code Quality2.6
CorpFin v254.2
CorpFin v2 - Exact Pages58.5
CorpFin v2 - Max Fitting Context51.9
CorpFin v2 - Shared Max Context52.3
CyberBench v1.1 Overall29.5
CyberBench v1.1 Patch33.9
CyberBench v1.1 PoC25.0
Excel Modeling Benchmark17.0
Excel Modeling Benchmark Comps14.3
Excel Modeling Benchmark Dataroom28.6
Excel Modeling Benchmark DCF11.2
Excel Modeling Benchmark LBO15.2
Excel Modeling Benchmark M&A19.5
Excel Modeling Benchmark Misc20.3
Excel Modeling Benchmark Operating9.8
Excel Modeling Benchmark Scratch17.4
Excel Modeling Benchmark Template16.0
Finance Agent v218.5
Finance Agent v2 - Adjustments14.0
Finance Agent v2 - All Pass10.3
Finance Agent v2 - Comparables24.3
Finance Agent v2 - Disclosure Analysis20.6
Finance Agent v2 - Earnings Analysis15.6
Finance Agent v2 - Financial Modeling4.6
Finance Agent v2 - General Qualitative Analysis33.0
Loading Atlas data…