NVIDIANemotronAI AgentsOpen ModelsBenchmarksLLMNemotron 3.5 Lightning

NVIDIA Nemotron 3.5 Lightning: Architecture, Benchmarks, and Hardware

NVIDIA Nemotron 3.5 Lightning is a 30B MoE model with 3B active parameters and up to 1M context. Review its published benchmark results and checkpoint-specific hardware guidance.

NVIDIA Nemotron 3.5 Lightning: Architecture, Benchmarks, and Hardware
Analyzed Foundation Models (2)Full LLM Leaderboard→
Nemotron 3 Ultra (550B A55B)
NVIDIA
37.0SCORE
DeepSeek-V4-Flash-0731
DeepSeek
46.1SCORE

NVIDIA’s Nemotron 3.5 Lightning model card describes a 30-billion-parameter model with 3 billion active parameters, using interleaved Mamba-2 and MoE layers with selected attention layers. It supports context up to one million tokens. NVIDIA publishes both BF16 and NVFP4 checkpoints and documents different deployment paths for each.

The “3B active” figure describes computation routing, not the full memory footprint. The full checkpoint stores 30B parameters, and serving also needs runtime buffers and context memory. NVIDIA’s model card lists single-GPU deployment for particular hardware/checkpoints, including one H100 for its documented NVFP4 deployment; that guidance should not be generalized to arbitrary precision, context size, or GPU.

Reported model details

ItemNVIDIA documentation
Total / active parameters30B total / 3B active
ArchitectureHybrid Mamba-2, MoE, and attention
ContextUp to 1M tokens
WeightsBF16 and NVFP4 checkpoints are published
LicenseNVIDIA OpenMDW 1.1
Release dateAugust 11, 2026

NVIDIA’s announcement claimed up to four times the output speed of similarly sized models. Treat “up to” as a vendor-reported maximum, not a throughput expectation for every serving setup. Tokens per second depend on hardware, precision, runtime, prompt length, batching, speculative decoding, and sampling settings.

Benchmark results and scope

The official NVFP4 model card reports benchmark results for both BF16 and NVFP4 checkpoints and notes that NVIDIA measured them with its NeMo Gym / NeMo Evaluator harness. It also says some scores may differ from vendor-reported results elsewhere. Selected published values are:

EvaluationBF16NVFP4
MMLU Pro81.9481.62
GPQA Diamond, no tools75.4475.57
SWE-bench Verified51.5652.80
SWE-bench Multilingual39.3336.47
Terminal-Bench 2.124.5823.46
PinchBench85.3783.43

NVIDIA publishes recipes with benchmark-specific prompts, containers, inference parameters, parsers, and scoring settings. These figures therefore describe a specified evaluation rather than a universal success rate. Scores from different benchmark versions or harnesses should not be compared as though they were measured under identical conditions.

Hardware and local deployment

Choose a checkpoint and runtime based on the hardware available, then follow NVIDIA’s deployment recipe for that combination. Its model card documents use with vLLM, TensorRT-LLM, SGLang, Ollama, llama.cpp, and LM Studio, and identifies specific validated hardware. Context size also changes memory needs: the model card notes that Ollama’s default context varies with available VRAM and that higher settings may offload work to the CPU.

Parameter counts alone cannot tell you whether a model fits in a GPU. Thirty billion parameters at one byte each would represent about 30 GB of raw weights before other memory needs; lower-precision formats can reduce weight storage, while larger context, runtime buffers, and KV state add their own requirements. Use the storage size and validated setup for the checkpoint you actually intend to run. A 3B active count does not make 3B of VRAM sufficient.

For coding-agent use, begin with a bounded repository task, observe tool and memory use, and compare results under a fixed prompt and evaluation procedure. A high SWE-bench score does not itself guarantee safe shell behavior or reliable fixes in a different codebase.

Takeaway

Nemotron 3.5 Lightning combines a sparse 30B model with a hybrid architecture and publishes multiple checkpoints and deployment recipes. The 3B active figure is not a VRAM estimate; use checkpoint-specific guidance. NVIDIA’s benchmark table is useful when read with its stated harness and should not be treated as a prediction for all hardware or coding-agent tasks.

Frequently Asked Questions

What is NVIDIA Nemotron 3.5 Lightning?

It is a 30-billion-parameter model with 3 billion active parameters, using a hybrid Mamba-2, MoE, and attention architecture. NVIDIA's model card lists context up to one million tokens.

How fast is Nemotron 3.5 Lightning?

NVIDIA's announcement claims up to four times the output speed of similarly sized models. This is a vendor claim; actual throughput depends on checkpoint, hardware, serving stack, batch size, and context.

What does NVIDIA report on SWE-bench Verified?

NVIDIA's official model card reports 51.56% for the BF16 checkpoint and 52.80% for the NVFP4 checkpoint, measured with NVIDIA's stated evaluation harness. The page notes results can differ from other vendors' reports.

Can a model with 3B active parameters run in 3B of GPU memory?

No. Active parameter count is not the total weight storage requirement. Nemotron has 30B total parameters, and required memory also depends on checkpoint precision, context, runtime, and serving setup.

Sources & disclosures

Source links for data and claims discussed in this article are listed below. Refer to each source for its current figures and methodology.

Share Article

Recommended Reading

Related by topic and recency.

All articles→
ComparisonsMuse Spark 1.3

Muse Spark 1.3 vs Gemini 3.8 Flash: Coding and Autonomous Agent Comparison

Compare what Meta documents for Muse Spark 1.3 with the checks needed to evaluate a Gemini Flash endpoint on matched agent tasks.

Sep 3, 2026 · 2 min read
ComparisonsDeepSeek V4 Pro

DeepSeek V4 Pro 0813 vs GLM-5.3: Context and API Pricing

Compare DeepSeek V4 Pro 0813 and GLM-5.3 context limits, current native API pricing, and a reproducible way to evaluate coding tasks.

Aug 17, 2026 · 2 min read
NewsGLM-5.3

GLM-5.3 in ZCode: Agentic Coding Integration and Evaluation

Z.ai describes GLM-5.3 as a post-training update to GLM-5.2 and says it rolled out to GLM Coding Plan users. Learn how to evaluate the model separately from the ZCode agent.

Aug 17, 2026 · 2 min read