Nemotron 3.5 Lightning

3.5 Lightning (30B-A3B) · Family: Nemotron

NVIDIA open reasoning LLM (30B total, 3B active) with a hybrid Mamba-2 + MoE + Attention architecture and up to 1M-token context.

✓ Active✓ Public access⚖ Open sourceLLMReasoning model📁 Nemotron

Context window

1M (256K na 1× H100)

tokens

Parameters

30B (3B active)

parameters

Release date

11 August 2026

🏢NVIDIAProducer

Access:APIDownloadHostedDeployment:💻 Local☁ Cloud📱 On-device

Overview

Nemotron 3.5 Lightning (NVIDIA-Nemotron-3.5-Lightning-30B-A3B) is an open large language model in NVIDIA's Nemotron family, trained by NVIDIA Corporation. It uses a hybrid Mixture-of-Experts architecture with interleaved Mamba-2 and MoE layers plus select Attention layers (NemotronHForCausalLM implementation). The model has 30B total parameters and 3B active parameters.

It was pre-trained on over 20T tokens with an NVFP4 recipe and includes Multi-Token Prediction (MTP) layers. It supports a configurable reasoning mode (enable_thinking) and tool calling. Context length reaches up to 1M tokens (256K on a single H100 in BF16). The model is released under the OpenMDW-1.1 license with open weights, training data and recipes. It is available in BF16 (reference, for customization) and NVFP4 (deployment-optimized) precision variants.

Classification

LLMReasoning model

Family: Nemotron

Access & deployment

APIDownloadHosted

LocalCloudOn-device

Weights: Open source

Key parameters

📏 Context: 1M (256K na 1× H100)

🧩 Parameters: 30B (3B active)

✓ Tools · ✓ Fine-tuning

📥 Input: text

Technical specification

Context window

1M (256K na 1× H100)

tokens

Parameters

30B (3B active)

parameters

License

OpenMDW-1.1

Hardware requirements

Single-GPU deployment: 1x H100 80GB (or A100 80GB) at 256K context in BF16. Full 1M window requires 8x H100, GB200 or B200. Supported architectures: NVIDIA Blackwell (GB200, B200), Hopper (H100, H200), Ampere (A100).

Features:✓ Tool use✓ Fine-tuning

Modalities

⬇ Input

text

⬆ Output

textcode

Capabilities and applications

Native model capabilities

Reasoning

The model's ability to reason logically and solve complex problems.

Category: reasoning

Coding

Generating, analysing and modifying code in many programming languages. Covers writing functions, debugging, refactoring, code review, and creating tests. Measured by benchmarks such as HumanEval and SWE-bench.

Category: coding

Tool use

The model's ability to call external functions, APIs and tools during a conversation: calculator, search engine, code editor, database. The model decides when and how to use a tool and interprets its result.

Category: planning

Long context

Support for large context windows — tens to hundreds of thousands (or millions) of input tokens. Enables analysis of entire codebases, long documents, and many parallel conversations without losing earlier information. GPT-5.1 supports 400,000 tokens.

Category: language

Instruction following

Precisely following instructions contained in the prompt: response format, length, style, constraints (e.g. 'reply in six words'). GPT-5.1 significantly improved this capability compared to GPT-5.

Category: language

Multilingual

Competence in many natural languages (from a few to over a hundred): understanding, generation, translation, and code-switching within a single conversation. Frontier models support a wide range of languages with comparable quality.

Category: language

Benchmark results

6 benchmarks

MMLU-Pro

accuracy · NeMo Gym harness

81.94%