trohrbaugh/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16-heretic

TEXT GENERATIONConcurrent Unit Cost:2Model Size:30BQuant:FP8Context Size:32kPublished:Aug 12, 2026License:otherArchitecture:Transformer0.0K Featherless Exclusive Cold

NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 is a 30 billion parameter Mixture-of-Experts (MoE) large language model developed by NVIDIA, featuring a hybrid Mamba-2, MoE, and Attention architecture with 3 billion active parameters. This BF16 full-precision release is primarily intended as a reference model for customization, including post-training, domain adaptation, and the creation of quantized variants. It supports a context length of up to 1 million tokens and excels as a foundation for specialized AI agents and research at full precision.

Loading preview...

Model Overview

NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 is a 30 billion parameter Mixture-of-Experts (MoE) large language model from NVIDIA, utilizing a hybrid architecture that combines Mamba-2, MoE, and Attention layers. It features 3 billion active parameters and supports an extensive context length of up to 1 million tokens, making it suitable for complex tasks.

Key Capabilities

  • Hybrid Architecture: Integrates Mamba-2, MoE, and Attention for efficient and powerful processing.
  • High Precision: Released in BF16 full-precision, serving as a robust reference for further development.
  • Extensive Context: Supports up to 1 million tokens, enabling deep contextual understanding and generation.
  • Multilingual Support: Primarily English and coding languages, with additional support for Spanish, French, German, Italian, and Japanese.
  • Configurable Reasoning: Offers an enable_thinking option in its chat template to control reasoning mode.
  • Speculative Decoding: Incorporates DSpark, DFlash, and Multi-Token Prediction (MTP) for faster text generation.

Ideal Use Cases

This model is best suited for developers focused on customization and post-training rather than direct production inference. It is designed for:

  • Supervised Fine-Tuning (SFT), Reinforcement Learning (RL), and distillation.
  • Domain adaptation and task-specific specialization.
  • Producing custom quantized variants (e.g., NVFP4, W4A16, GGUF).
  • Full-precision research and evaluation workflows.