What is Nemotron 3.5 Lightning?
Nemotron 3.5 Lightning is the smallest model in NVIDIA's Nemotron 3 family. It mixes Mamba-2, mixture-of-experts and attention layers, so only 3B of its 30B parameters work on each token, and it ships with draft models for speculative decoding. It reads and writes text across up to 1M tokens, with reasoning you can switch on or off.
NVIDIA pitches it for the execution layer of agents: tool calls, checks and sub-agent jobs, while a bigger model does the planning. It is not the most accurate model of its size; NVIDIA's own table puts Qwen 3.6 35B A3B ahead on most tests.
What changed from Nemotron 3 Nano?
Mostly accuracy on agent and coding tests, plus faster decoding, at the same 30B size with 3B active.
- Coding
- 51.6 on SWE-bench Verified and 24.6 on Terminal-Bench 2.1, up from 34.1 and 8.3 in NVIDIA's table.
- Web research
- 37.0 on BrowseComp, up from 13.7.
- Long context
- 52.0 on AA-LCR, up from 32.8.
- Speed
- Multi-token prediction is trained in, as in Nemotron 3 Super and Ultra, and two draft models ship with it.
Who is Nemotron 3.5 Lightning best for?
Nemotron 3.5 Lightning is best for sub-agents and tool calls, local agents on one GPU, fine-tuning for a single job, and routing setups that pair it with a bigger planning model.
Sub-agents and tool calls
NVIDIA trained it for agent harnesses such as OpenClaw and Hermes Agent and built it for high volumes of short, repeated calls.
Local agents on one GPU
The 4-bit build runs on a single DGX Spark or H100 and on a GeForce RTX 5090, and Ollama, LM Studio and llama.cpp support it.
Fine-tuning for one job
Weights, data and recipes are open, and NVIDIA supports LoRA, full fine-tuning and reinforcement learning with its NeMo tools.
Model routing setups
NVIDIA's NeMo Switchyard can send execution steps to Lightning and planning to a larger frontier model.
What's included with Nemotron 3.5 Lightning?
Built-in speculative decoding, NVFP4 and BF16 checkpoints, reasoning you can switch off, tool calling, six languages and open training data.
- Built-in speculative decoding
- Multi-token prediction is trained in, and DSpark and DFlash draft models ship with it for different serving loads.
- NVFP4 and BF16 checkpoints
- The 4-bit NVFP4 build runs on Blackwell, Hopper and Ampere GPUs; BF16 is the full-precision base for fine-tuning.
- Reasoning on or off
- Thinking is on by default and can be turned off per request through the chat template.
- Tool calling and structured output
- Function calling and structured output are supported, including on NVIDIA's hosted endpoint.
- Six languages
- English and code come first, with Spanish, French, German, Italian and Japanese also supported.
- Open training data
- NVIDIA released the training data and recipes with the weights, plus an open agentic reinforcement learning dataset.
How much does Nemotron 3.5 Lightning cost?
Nemotron 3.5 Lightning is free: the weights download at no cost with commercial use allowed and NVIDIA offers a free trial API, while NVIDIA's packaged NIM container needs a paid AI Enterprise licence in production.
- OpenMDW-1.1 licence, commercial use allowed
- BF16 and NVFP4 checkpoints on Hugging Face
- You pay only for your own hardware
- For prototyping and testing, not production
- Rate limits vary by model and load
- Inputs and outputs are logged to improve NVIDIA products
- Needed only to run NVIDIA's NIM container in production
- 90-day free evaluation first
- Includes NVIDIA enterprise support
NVIDIA lists no per-token price for Lightning. NVIDIA AI Enterprise also comes as a perpetual licence with five years of support at $22,500 per GPU, or at $1 per GPU per hour plus instance costs on cloud marketplaces; qualified education and research institutions and NVIDIA Inception or Connect members pay $1,125 per GPU per year. Hosted providers such as Together AI, Fireworks AI, DeepInfra and Baseten set their own prices, and NVIDIA links a free version on OpenRouter. Full pricing on build.nvidia.com. We checked these prices on .
How do readers rate Nemotron 3.5 Lightning?
No one has rated Nemotron 3.5 Lightning yet. Be the first.
Sign in to rate. One rating per account.
Nemotron 3.5 Lightning pros and cons
It is fast, fully open and good at calling tools, but it writes weak code on its own and tends to overthink.
Pros
- Fast: NVIDIA claims up to 4x the output speed of similar-sized models, and users report about 50 to 100 tokens a second on MacBooks and about 150 on a DGX Spark.
- Users running it in Hermes Agent and OpenCode say it calls tools reliably, in some tests more readily than Gemma 4.
- Fully open: weights, training data and recipes come under OpenMDW-1.1 with commercial use allowed, which commenters say few models this capable offer.
- A clear step up from Nemotron 3 Nano: NVIDIA's table shows 51.6 against 34.1 on SWE-bench Verified, and users who ran both agree it is better.
Cons
- Weak at writing code and apps on its own: several users called its output poor, and NVIDIA's own table puts Qwen 3.6 35B A3B at 70.1 on SWE-bench Verified against 51.6.
- Tends to overthink: several testers saw long, second-guessing reasoning traces before fairly simple answers.
- Some users found it ignores instructions or declares a task done too early outside simple, structured jobs.
- Text only, and the full 1M context needs big hardware: the BF16 version on one H100 tops out around 256K tokens.
Is Nemotron 3.5 Lightning open source?
Yes. Nemotron 3.5 Lightning's code is public under the OpenMDW 1.1 licence, so you can run it on your own server.
What are the best Nemotron 3.5 Lightning alternatives?
Nemotron 3.5 Lightning is an open source alternative to Gemini 3.5 Flash-Lite and GPT-6 Luna. DeepSeek V4.1 Flash, 1-bit Bonsai, Bonsai 27B and Mellum2 do the same job and are open source too.
Gemini 3.5 Flash-LiteFreeGoogle's low cost Gemini 3.5 model for high volume tasks and subagents.
GPT-6 LunaFreeOpenAI's lowest cost GPT-6 model for focused, high volume tasks.
DeepSeek V4.1 FlashOpen SourceDeepSeek's small, efficient open model.
1-bit BonsaiOpen SourcePrismML's 1-bit language models in 8B, 4B and 1.7B sizes that run on phones and laptops.
Bonsai 27BOpen SourcePrismML's 1-bit and ternary 27B models that run locally on modest hardware.
Mellum2Open SourceJetBrains' open 12B mixture-of-experts model for fast routing, RAG and sub-agent steps in coding systems.
Questions about Nemotron 3.5 Lightning
What is Nemotron 3.5 Lightning?
Nemotron 3.5 Lightning is NVIDIA's free, open 30B mixture-of-experts model with 3B active parameters, built to run the quick, repetitive tool-calling steps inside long-running AI agents.
Who is Nemotron 3.5 Lightning for?
Nemotron 3.5 Lightning is best for sub-agents and tool calls, local agents on one GPU, fine-tuning for a single job, and routing setups that pair it with a bigger planning model.
Is Nemotron 3.5 Lightning free?
Yes. Nemotron 3.5 Lightning has a free Open weights plan: openMDW-1.1 licence, commercial use allowed.
How much does Nemotron 3.5 Lightning cost?
Nemotron 3.5 Lightning's paid plans are NVIDIA AI Enterprise at $4,500 per GPU per year.
Is Nemotron 3.5 Lightning open source?
Yes. Nemotron 3.5 Lightning's code is public under the OpenMDW 1.1 licence, so you can run it on your own server.
What are the best Nemotron 3.5 Lightning alternatives?
The closest alternatives on Recaply are Gemini 3.5 Flash-Lite, GPT-6 Luna, DeepSeek V4.1 Flash, 1-bit Bonsai, Bonsai 27B and Mellum2.
Where can I try Nemotron 3.5 Lightning?
On its official website, developer.nvidia.com.
Can I use Nemotron 3.5 Lightning commercially?
Yes. NVIDIA releases Nemotron 3.5 Lightning under the OpenMDW-1.1 licence and says it is ready for commercial use. The paid NVIDIA AI Enterprise licence applies to running NVIDIA's packaged NIM container in production.
What hardware do I need to run Nemotron 3.5 Lightning?
The 4-bit NVFP4 build runs on a single DGX Spark or H100, and NVIDIA supports Blackwell, Hopper and Ampere GPUs, including the GeForce RTX 5090. Ollama and LM Studio run it locally; on one H100, the BF16 version is limited to about 256K tokens of context.
Is Nemotron 3.5 Lightning better than Qwen 3.6 35B?
Not on accuracy. In NVIDIA's own table, Qwen 3.6 35B A3B scores higher on most tests, including SWE-bench Verified (70.1 against 51.6). NVIDIA's case for Lightning is speed: it says Lightning finishes 10,000 PinchBench tasks 30% faster at similar accuracy.
Does Nemotron 3.5 Lightning accept images?
No. Nemotron 3.5 Lightning takes text in and gives text out. For image, video or audio input from the same family, NVIDIA offers Nemotron 3 Nano Omni instead.
Our verdict
Nemotron 3.5 Lightning is a good pick for the fast, repetitive steps inside an agent, not for writing your code.
It is fully open, runs on one GPU, and users report high speed and reliable tool calls, but NVIDIA's own table shows Qwen 3.6 35B A3B ahead on most accuracy tests.
- Skip it for
- Coding agents that build whole features, and anything that needs image input.
- Check first
- Test it on your own tool-calling workload, with reasoning on and off.
NVIDIA Nemotron 3 Ultra
Nemotron 3
Nemotron 3 Nano Omni
Nemotron 3 Super
Reflection Beam
DiffusionGemma