# NVIDIA BF16 reasoning evaluation, not the NVFP4 checkpoint. Benchmark-specific recipes:
# https://github.com/NVIDIA-NeMo/Gym/tree/main/nemotron_recipes/lightning-3.5/reproducibility.md
name = "Nemotron 3.5 Lightning 30B A3B"
description = "Fast NVIDIA Nemotron MoE for reliable agentic tasks across enterprise workloads"
family = "nemotron"
release_date = "2026-08-11"
last_updated = "2026-08-11"
attachment = false
reasoning = true
temperature = true
tool_call = true
structured_output = true
open_weights = true

[limit]
context = 262_144
output = 262_144

[modalities]
input = ["text"]
output = ["text"]

[[benchmarks]]
name = "MMLU-Pro"
score = 81.94
metric = "score"
variant = "BF16; reasoning"
harness = "NeMo Gym / NeMo Evaluator SDK"
source = "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16"

[[benchmarks]]
name = "AA-Omniscience"
score = 17.5
metric = "score"
variant = "BF16; reasoning"
harness = "NeMo Gym / NeMo Evaluator SDK"
source = "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16"

[[benchmarks]]
name = "SciCode"
score = 32.6
metric = "score"
variant = "BF16; reasoning"
harness = "NeMo Gym / NeMo Evaluator SDK"
source = "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16"

[[benchmarks]]
name = "PinchBench"
score = 85.37
metric = "score"
variant = "BF16; reasoning"
harness = "NeMo Gym / NeMo Evaluator SDK"
source = "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16"

[[benchmarks]]
name = "BrowseComp"
score = 36.97
metric = "score"
variant = "BF16; reasoning"
harness = "NeMo Gym / NeMo Evaluator SDK"
source = "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16"

[[benchmarks]]
name = "AA-LCR"
score = 52
metric = "score"
variant = "BF16; reasoning"
harness = "NeMo Gym / NeMo Evaluator SDK"
source = "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16"

[[benchmarks]]
name = "GPQA Diamond"
score = 75.44
metric = "accuracy"
variant = "BF16; reasoning; no tools"
source = "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16"

[[benchmarks]]
name = "Humanity's Last Exam"
score = 11.72
metric = "accuracy"
variant = "BF16; reasoning; no tools"
dataset = "text-only"
source = "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16"

[[benchmarks]]
name = "SWE-Bench Verified"
score = 51.56
metric = "accuracy"
variant = "BF16; reasoning"
harness = "NeMo Evaluator"
source = "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16"

[[benchmarks]]
name = "SWE-Bench Multilingual"
score = 39.33
metric = "accuracy"
variant = "BF16; reasoning"
harness = "NeMo Evaluator"
source = "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16"

[[benchmarks]]
name = "Terminal-Bench"
score = 24.58
metric = "accuracy"
variant = "BF16; reasoning"
harness = "NeMo Evaluator"
version = "2.1"
source = "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16"

[[benchmarks]]
name = "tau3-bench"
score = 9.28
metric = "score"
variant = "BF16; reasoning; banking"
source = "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16"

[[benchmarks]]
name = "GDPval-AA"
score = 832
metric = "Elo"
variant = "BF16; reasoning"
version = "2"
source = "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16"

[[benchmarks]]
name = "IFBench"
score = 71.88
metric = "accuracy"
variant = "BF16; reasoning; loose"
source = "https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16"