Are the weights open and can I run this locally?

Yes, the BF16 weights are on Hugging Face under the OpenMDW License v1.1 and are cleared for commercial use. A single H100 80GB or A100 80GB handles it, though for consumer hardware you will want the NVFP4 or GGUF variants NVIDIA references separately.

What does it cost and what is the context window?

Via API it is $0.07 per million input tokens and $0.20 per million output tokens. Context is 256K tokens on a single H100; the architecture supports up to 1M tokens on multi-GPU Blackwell setups.

How does this BF16 release differ from the NVFP4 release?

The BF16 release is the full-precision reference checkpoint, intended for post-training work such as SFT, RL, distillation, and producing your own quantized variants. The NVFP4 release is the one NVIDIA recommends for optimized inference and low-latency deployment.

▲

59.6%

this week

Nemotron 3.5 Lightning

8980°

#llm #nvidia #reasoning #agents #local #long context

This month

Run this model

Offers / Ads

OpenRouter Terminal / API

Pay per token

Chat with this model in the browser or call it from your code with one API key.

NVIDIA | 262,144 token context | $0.07/M input, $0.20/M output. NVIDIA Nemotron 3.5 Lightning is an open mixture-of-experts model from NVIDIA, with 3B active parameters out of 30B total. It is suited for high-throughput agentic workloads and specialized tasks that benefit from domain-specific customization.

Guides for this model

Ollama vs LM Studio (2026): which local AI app to use, and which top models each one runs

Best local LLM by use case (October 2026): the right open model for every job

Open weights vs open source vs API-only: what you are actually allowed to do

What is a GGUF file? Quantized models explained (Q4, Q8, fp8, NF4)

How much VRAM do I actually need for local AI?

How to Install Ollama and Run a Local AI (Windows, Mac, Linux)

All guides

Captain's notes

Written by Captain, the house editor

Notes from 2026-10-13

Nemotron 3.5 Lightning is a 30B mixture-of-experts model from NVIDIA with only 3B parameters active at inference, built on a hybrid Mamba-2, MoE, and Attention architecture. The BF16 release here is the reference checkpoint — the right starting point for fine-tuning, distillation, or building your own quantized variants, not the fastest path to production. Open weights, commercial use permitted, and a 256K context on a single H100.

Known for

MoE efficiency: 30B total parameters, only 3B active per forward pass
Hybrid architecture mixing Mamba-2, MoE layers, and Attention
Configurable reasoning mode — toggle thinking on or off at inference
Designed as a fine-tuning and distillation base, not just a chat endpoint
Supports six languages including English, Spanish, French, German, Italian, Japanese

Watch out for

This BF16 release is for customization work; for low-latency serving, NVIDIA points you to the NVFP4 variant instead.
Single H100 deployment is validated to 256K tokens; reaching 1M requires multi-GPU or Blackwell hardware.
Weights are open but under the OpenMDW License Agreement v1.1 — not a standard Apache or MIT licence; check the terms.

Best for

Prompts to try

Toggle reasoning on a problem

You are solving a logistics problem. A warehouse ships orders in batches of 12. It has 500 units of product A and receives 75 new orders each day. Each order requests 3 units. How many full days until stock runs out, and on which day does the last complete batch ship? Show your reasoning step by step.

Domain fine-tune prep

I am preparing a supervised fine-tuning dataset for a legal document review assistant. Write 5 high-quality instruction-response pairs where the instruction asks the model to identify a specific clause type in a contract excerpt, and the response identifies the clause, quotes the relevant text, and explains its significance in plain English.

Agentic task decomposition

You are an agent with access to a web search tool, a code execution tool, and a file-write tool. A user wants a weekly report on competitor pricing for three SaaS products. Break this into a concrete sequence of tool calls and sub-tasks, specifying what each call fetches or computes and how the outputs feed into the final report.

Compare with

Nemotron 3 Ultra
Open-weights model
262,144 token context · $0.60/M input, $2.40/M output
Nemotron 3 Super
Open-weights model
262,144 token context · $0.08/M input, $0.45/M output
NVIDIA Nemotron 3 Ultra 550B A55B
Open-weights model
262,144 token context · $0.50/M input, $2.20/M output
Nemotron 3.5 Content Safety
Open-weights model
131,072 token context · $0.20/M input, $0.20/M output

More text models from nvidia

Questions people ask

Are the weights open and can I run this locally?: Yes, the BF16 weights are on Hugging Face under the OpenMDW License v1.1 and are cleared for commercial use. A single H100 80GB or A100 80GB handles it, though for consumer hardware you will want the NVFP4 or GGUF variants NVIDIA references separately.
What does it cost and what is the context window?: Via API it is $0.07 per million input tokens and $0.20 per million output tokens. Context is 256K tokens on a single H100; the architecture supports up to 1M tokens on multi-GPU Blackwell setups.
How does this BF16 release differ from the NVFP4 release?: The BF16 release is the full-precision reference checkpoint, intended for post-training work such as SFT, RL, distillation, and producing your own quantized variants. The NVFP4 release is the one NVIDIA recommends for optimized inference and low-latency deployment.

Captain's verdict
A serious fine-tuning base with a clever MoE architecture — reach for it when you need to customize, not when you just need a chat endpoint.

Model card on Hugging Face

Captain writes these notes from the model's own description and model card. Prices and context sizes are the provider's at the time of writing and can change.

Ranking badge

Updated Oct 22, 2026

Made this model? Show this badge on your page.

Paste it into your Hugging Face card or Civitai description.

Change colors / personalize / get code

Also ranked: #54 Hottest AI model · #29 Hottest Text / LLM model · #29 Hottest Text / LLM text model

Not this creator?

Get your own badge

Nemotron 3.5 Lightning Popularity Trend

Heat score since 11 Aug 2026

Heat score

8,980°

This week

▲

59.6%

8,919° then, 8,980° now

This month

▲

1,079%

8,656° then, 8,980° now

Past year

n/a

Wasn't released/tracked

7,000°

Floor

The scale runs from 0° to 9,999°. Everything below 7,000° is kept for models no source reports on yet, which carry a stand-in number there so lists stay complete. 7,000° is where measured data begins: a model with real data but almost no downloads sits at the floor, and the measurements fill the degrees above it. The floor says nothing about the model; only the degrees above it do, and they are what separates one model from another.

+1,656°

OpenRouter tokens

3,296,276,417,858 tokens this model processed through OpenRouter in the last 30 days, prompts and answers together. Every 10,000,000 tokens count as one download in the download term, 329,627 in all; the last day's tokens join the adoption rate. Source: OpenRouter (openrouter.ai/rankings).

+324°

Adoption rate

About 21,182 downloads a day, measured between two readings of the sources, never guessed from a lifetime average. Speed weighs a quarter of the download term, so a fresh surge shows without letting a one-day spike outrank years of use.

Show detailed breakdown

Every number on this site can be checked. This model's heat score is 8,980°, and here it is built one ingredient at a time, with the running total after each: the floor, the degrees each source adds, the bonus for the kind of model and what the age of its weights takes away. The colors are the bands of the chart above. The glossary at the end defines every term.

1
Start at the floor.
Why 7,000° and not zero: the scale runs from 0° to 9,999°, and the part below 7,000° is reserved for models no source reports on yet, which carry a stand-in number there so the lists stay complete and never outrank a model with data. 7,000° is where measured data begins. A model with real data but almost no downloads sits here; everything that follows is measurement, and the room above the floor holds ten tenfold steps in downloads at 300° each. The floor is the same for every model, so it decides nothing; the degrees above it do.
7,000°
running total
2
Add the downloads.
329,627 download-equivalents in all, from OpenRouter tokens. Ten times the downloads adds the same 300°, so size matters but never runs away with the score.
+1,656°
8,656°
running total
3
Add the adoption rate.
About 21,182 downloads a day right now, measured between two readings, at a quarter of the weight.
+324°
8,980°
running total
=
The heat score.
Where this model stands on every list of this site. Nothing else adjusts it.
8,980°

Glossary: the terms used here

Heat score: One number for how much a model is used and wanted right now, from 0° to 9,999°. Every list on this site that says hottest is sorted by it and nothing else.
Floor: The 7,000° every model with real data starts from. The scale runs from 0° to 9,999°; the part below the floor is reserved for stand-ins (models without data), so that a model with data always ranks above one without. The floor is the same for everyone and decides nothing; the degrees above it are the measurement.
Download-equivalents: Everything the score treats as a download, added up: downloads themselves, each torrent seeder as 1,000, each AI Horde image and each run on a generation site as one.
Log scale: Counting by tens, not by ones: going from 1,000 to 10,000 downloads adds the same degrees as going from 10,000 to 100,000. It keeps giants from crowding out everything else.
Likes: Thumbs up and likes from every source, worth half a download each on the log scale.
Adoption rate: How many downloads a day a model gets right now, from the difference between two readings of a source. A quarter of the weight of the download term.
Measured and estimated: Measured means two readings of a source exist and the difference is real. Estimated means only today's total is known and the past is spread over the model's life; the site says which one it is showing.
Seeder: Someone who keeps a copy of the weights online for others to download over a torrent, or a hosted mirror doing the same. Each one counts as 1,000 downloads.
AI Horde: A volunteer network that generates images for free on donated computers. The number of images it has made with a model is public, and each counts as one download.
Generation sites: Sites that host their own copies of a model and let people generate with it, such as PixAI, SeaArt and Yodayo. Their downloads count like Civitai's and each generation run as one download.
OpenRouter tokens: A text model served through an API, such as Claude or GPT, has no weights to download, so its usage is measured in tokens: the pieces of text it read and wrote through OpenRouter, a gateway many apps use, over the last 30 days. Every 10,000,000 tokens count as one download. Source: OpenRouter (openrouter.ai/rankings), CC BY 4.0.
Full model (checkpoint): The weights themselves, the file that generates on its own. It gets a bonus because add-ons depend on it.
LoRA and embedding: Small add-on files that steer someone else's weights toward a style, a character or a concept. They cannot run alone.
Family and weights: The base a model is built on, such as SD 1.5, SDXL, Flux or Wan. Every model inherits the release date of its family's weights.
Age: Degrees taken off for how long ago the family's weights came out: 150° a year, up to 750°. It keeps old downloads that had years to pile up from freezing the top of the list.
Family boost: A small, bounded nudge for LoRAs of a family this site is pushing forward right now, always listed in the breakdown.
Stand-in score: A stable number below the floor given to a model no source reports on yet, so lists stay complete. It never ranks above a model with data.
New! (grace period): A model less than 7 days old shows New! instead of a score and holds no place in the rankings. Its score is computed from day one, but a model one day old grows faster than anything, so without the wait the week's #1 would always be whatever came out yesterday rather than what held up for a week.
Trend: How a model's growth over a period compares with the typical model's growth over the same period. Faster than typical reads as a rise, slower as a fall.
Typical model: The middle of the catalog: half the models with a figure grew faster over the period, half slower.

How this model's popularity has moved against the rest of the catalog: its growth in downloads, likes and usage over each period, measured against the typical model's growth over the same period, with its heat score at the time. Above the typical growth reads as a rise, below it as a fall. The past is reconstructed from today's totals, the model's first day and its measured download speed where two readings exist; a model that did not exist at the start of a period shows n/a.

Similar Models

NVIDIA Nemotron 3 Super 120B A12B

TEXT / LLM

AI Models for Image, Video & Text Generation

Nemotron 3.5 Lightning is a text model for the Text / LLM family listed on ModelTrends.ai, a read-only catalog of open source AI image, video and text models. Compare it with other Text / LLM models and check its heat score to see how it is trending.

More Text / LLM models llm nvidia reasoning agents local long context multilingual

Browse by Model Family