NVIDIA Nemotron 3.5 Lightning 30B-A3B
Nemotron 3.5 Lightning 30B-A3B is a small hybrid Mamba-2/Transformer mixture-of-experts model that NVIDIA released in August 2026. It has 30B total and 3B active parameters, a context of up to 1M tokens and thinking that is on by default but can be turned off. NVIDIA says it deploys on a single 80 GB H100 or A100.
Specifications
- Context window
- 1Mabout 750,000 words, or 1,500 pages
- Max output
- Not published
- Input price
- $0.05per 1M tokens
- Output price
- $0.20per 1M tokens
- Cached input
- $0.01per 1M tokens
- Weights
- OpenOpenMDW-1.1
- Parameters
- 30B total, 3B activeMoE
- Input types
- Text
- Output types
- Text
- Reasoning mode
- Yes
- Tool calling
- Yes
- Released
- Aug 11, 2026Current
Fireworks standard serverless price; DeepInfra lists $0.06 / $0.03 cached / $0.16. Official context is up to 1M; the model card's single-H100 example uses 256K. The Hugging Face BF16 checkpoint stores 31.6B parameters including multi-token-prediction layers.
What it costs in practice
| Workload | Monthly cost at list price |
|---|---|
| Support chatbot30,000 conversations a month, 2,000 input and 400 output tokens each | $5.40 |
| Document summaries2,000 documents a month, 12,000 input and 800 output tokens each | $1.52 |
| Coding agent1,000 tasks a month, 60,000 input and 6,000 output tokens each | $4.20 |
Standard rates, no caching or batch discounts. Run your own numbers.
Strengths
- Single 80 GB GPU deployment
- Very low hosted prices
- Hybrid design with only 6 attention layers keeps KV cache small
Trade-offs to weigh
- Text-only input
- NVIDIA recommends the NVFP4 build for production; BF16 is aimed at customization
Where it fits
NVIDIA positions NVIDIA Nemotron 3.5 Lightning 30B-A3B for:
Where you can use it
- Hugging Face
- Ollama
- Fireworks
- DeepInfra
- build.nvidia.com
Run it yourself
Released under OpenMDW-1.1. Approximate memory needed just to load the weights:
| Precision | Memory for weights |
|---|---|
| 16-bit (BF16/FP16) | 60 GB |
| 8-bit (FP8/INT8) | 30 GB |
| 4-bit (Q4_K_M, about 4.85 bits) | 18 GB |
This is a mixture-of-experts model: all 30B total parameters must sit in memory, but only about 3B are used per token, so it runs faster than a dense model of the same size.
Add memory for the KV cache (it grows with context length and concurrent users) and 1-2 GB of runtime overhead. Size it precisely in the VRAM calculator.
Alternatives to consider
| Model | Tier | Context | Input / output per 1M | |
|---|---|---|---|---|
| Ministral 3 3BMistral AI | Compact | 256K | $0.10 / $0.10 | Compare |
| Granite 4.2 8BIBM | Compact | 128K | $0.06 / $0.25 | Compare |
| gpt-oss-20bOpenAI | Compact | 128K | $0.075 / $0.30 | Compare |
| Gemma 4 26B A4BGoogle | Compact | 256K | $0.07 / $0.34 | Compare |
Frequently asked questions
How much does NVIDIA Nemotron 3.5 Lightning 30B-A3B cost?
NVIDIA Nemotron 3.5 Lightning 30B-A3B lists at $0.05 per million input tokens and $0.20 per million output tokens, with cached input at $0.01 (checked October 10, 2026). Fireworks standard serverless price; DeepInfra lists $0.06 / $0.03 cached / $0.16. Official context is up to 1M; the model card's single-H100 example uses 256K. The Hugging Face BF16 checkpoint stores 31.6B parameters including multi-token-prediction layers. Confirm current rates on the provider's pricing page before you commit.
What is the context window of NVIDIA Nemotron 3.5 Lightning 30B-A3B?
NVIDIA Nemotron 3.5 Lightning 30B-A3B accepts up to 1,000,000 tokens of context (about 750,000 words, or 1,500 pages). Long prompts cost more and can be slower, so most apps send only what the task needs.
Is NVIDIA Nemotron 3.5 Lightning 30B-A3B open source?
NVIDIA Nemotron 3.5 Lightning 30B-A3B is an open-weight model released under the OpenMDW-1.1 license. You can download the weights from Hugging Face (nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16) and run or fine-tune them yourself, within the license terms. Open weights isn't the same as fully open source: training data and code may not be published.
Can I run NVIDIA Nemotron 3.5 Lightning 30B-A3B locally?
Yes, if you have enough memory. The weights alone need about 18 GB at 4-bit quantization and about 60 GB at 16-bit, plus room for the KV cache. Use the VRAM calculator on LLModel.com to check a specific GPU or Mac.
What is NVIDIA Nemotron 3.5 Lightning 30B-A3B best for?
NVIDIA positions NVIDIA Nemotron 3.5 Lightning 30B-A3B for low latency, high volume, agents, self-hosting. It accepts text as input, has a reasoning (thinking) mode and supports tool calling. Test it on a sample of your own tasks before you choose it for production.
Sources
- huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16
- docs.fireworks.ai/serverless/pricing
- huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16/raw/main/config.json
- huggingface.co/api/models/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16
- deepinfra.com/pricing
- ollama.com/library
Facts checked against these pages on October 10, 2026. Spot an error? Tell us.
Still choosing a model?
Send us your use case and constraints. We'll come back with a shortlist and an evaluation plan you can run on your own data.
- Free and vendor-neutral: no model provider pays for placement.
- Your details go only to partners you agree to.