NVIDIA Nemotron 3.5 Lightning 30B-A3B Model Card
by NVIDIA
A 15,000-word open recipe: hybrid Mamba-2 MoE architecture, training data, RL stages and serving.
Overview
This is the BF16 model card for NVIDIA Nemotron 3.5 Lightning 30B-A3B, released 11 August 2026, and at over 15,000 words with 100+ tables it functions as a technical report rather than a download page. The architecture is a hybrid mixture of experts: interleaved Mamba-2 and MoE layers with select attention layers, 30B total parameters with 3B active per token, supporting up to a one-million-token context — the card states plainly that the practical ceiling is 256K on a single H100 for memory reasons. Training ran to over 20 trillion tokens on Megatron-LM in four stages: pretraining with an NVFP4 recipe over synthetic and crawled data, continued pretraining to add multi-token-prediction layers, supervised fine-tuning on code, maths, science and instruction data, and reinforcement learning with GRPO across multiple environments using NeMo RL and NeMo Gym. What makes it a learning resource rather than marketing is that NVIDIA published the data alongside the weights under OpenMDW-1.1 — Nemotron-CC at 9.1T tokens, roughly 2.1T tokens of code, Nemotron-Post-Training v1 and v2, and the Nemotron-RL Agentic Terminal Pivot dataset — and documents the exact customisation path through NeMo Automodel and Megatron Bridge. Reported results include MMLU Pro 81.94, GPQA Diamond 75.44, SWE-bench Verified 51.56 and PinchBench 85.37. Deployment recipes cover vLLM and SGLang on H100, GB200 and B200, with three speculative decoding options: MTP, DSpark and DFlash.
At a Glance
- Topic
- Models
- Level
- Advanced
- Format
- Documentation
- Cost
- Free
- Duration
- ~2 hours — a 15,000+ word card with 100+ dataset and benchmark tables
- Provider
- NVIDIA
- Hands-on
- Yes — code/exercises
- Certificate
- None
What You’ll Learn
- ✓How interleaved Mamba-2, MoE and attention layers combine into one hybrid architecture
- ✓Why 3B active of 30B total parameters changes serving cost and latency arithmetic
- ✓The four-stage recipe: NVFP4 pretraining, MTP continued pretraining, SFT, then GRPO
- ✓Which open datasets went into training, including Nemotron-CC at 9.1 trillion tokens
- ✓How to deploy on vLLM or SGLang using the published H100, GB200 and B200 recipes
- ✓When to choose MTP, DSpark or DFlash speculative decoding for your concurrency profile
- ✓What the OpenMDW-1.1 licence permits for SFT, RL, distillation and quantised derivatives
Highlights
- •Released 11 August 2026 with weights, training data and training recipes published together
- •Hit 262 points and 139 comments on Hacker News the day it launched
- •Documents a 1M-token context and then states the real 256K single-H100 limit honestly
- •Names the exact customisation path: NeMo Automodel, Megatron Bridge, NeMo RL and NeMo Gym
- •Positioned for post-training and derivatives rather than drop-in production use, and says so
Who It’s For
Best For
- ✓Engineers evaluating open-weight models for long-running agent execution
- ✓Teams post-training an open model on their own domain data with SFT or RL
- ✓Inference engineers tuning speculative decoding and serving throughput
Prerequisites
- •Familiarity with transformer and mixture-of-experts architectures
- •Experience serving models with vLLM, SGLang or TensorRT-LLM
- •Comfort reading benchmark tables and open-weight licence terms
FAQ
What is NVIDIA Nemotron 3.5 Lightning 30B-A3B Model Card?
The BF16 model card for NVIDIA's Nemotron 3.5 Lightning, a 30B mixture-of-experts model with 3B active parameters released in August 2026. It reads as a technical report rather than a download page, and is for engineers evaluating open models for agent execution or planning to post-train one. You learn the hybrid Mamba-2 architecture, the four-stage training recipe, exactly which open datasets went in, and how to serve it on vLLM or SGLang.
Is NVIDIA Nemotron 3.5 Lightning 30B-A3B Model Card free?
NVIDIA Nemotron 3.5 Lightning 30B-A3B Model Card is free to access.
What level is NVIDIA Nemotron 3.5 Lightning 30B-A3B Model Card for?
NVIDIA Nemotron 3.5 Lightning 30B-A3B Model Card is aimed at a advanced audience. Recommended background: Familiarity with transformer and mixture-of-experts architectures, Experience serving models with vLLM, SGLang or TensorRT-LLM, Comfort reading benchmark tables and open-weight licence terms.
How long does NVIDIA Nemotron 3.5 Lightning 30B-A3B Model Card take?
Expect roughly ~2 hours — a 15,000+ word card with 100+ dataset and benchmark tables. Most learners work through it at their own pace.
What will I learn from NVIDIA Nemotron 3.5 Lightning 30B-A3B Model Card?
You'll learn: How interleaved Mamba-2, MoE and attention layers combine into one hybrid architecture; Why 3B active of 30B total parameters changes serving cost and latency arithmetic; The four-stage recipe: NVFP4 pretraining, MTP continued pretraining, SFT, then GRPO; Which open datasets went into training, including Nemotron-CC at 9.1 trillion tokens; How to deploy on vLLM or SGLang using the published H100, GB200 and B200 recipes; When to choose MTP, DSpark or DFlash speculative decoding for your concurrency profile; What the OpenMDW-1.1 licence permits for SFT, RL, distillation and quantised derivatives.
Topics
Sources
This page was written from 3 sources, 2 on domains other than huggingface.co.