NVIDIA Nemotron Models: Ultra, Super, Lightning and Bolt

NVIDIA's Nemotron 3 models are open-weights language models built on a hybrid of Mamba-2, mixture-of-experts and attention layers. The line runs from Lightning (30B total, 3B active) through Super (120B, 12B active) to Ultra (550B, 55B active, 1M-token context). Licences are OpenMDW 1.1 or NVIDIA's own Nemotron Open Model License, both allowing commercial use. Checked on 31 October 2026.

The Nemotron 3 lineup

The hub also lists Nemotron 3 Nano Omni (30B-A3B with reasoning), a transcription model, embedding and reranking models, a document parser and domain models. We cover only the language models here. The release dates come from each card, except Nano and Bolt, where we used the hub's creation dates.

The naming has a pattern. Ultra, Super and Nano are the size tiers of the main Nemotron 3 line. The 3.5 releases, Lightning and Bolt, are newer variants aimed at specific jobs: Lightning for customising a fast small model, Bolt for running agent work on a single local GPU.

Maker, licence and open weights

NVIDIA publishes the weights on Hugging Face and describes them as ready for commercial use. Two licences cover the family.

The Nemotron licence is short and permissive. It has no revenue threshold, no usage ban list and no user-count limit in the text we read, which makes it closer to Apache 2.0 than to the custom licences of some other open-weights models. The patent-litigation termination clause is standard in that style. If you are comparing licences across models, the table on the open source LLM page is a good companion.

The shift from NVIDIA's own licence on the earlier models to OpenMDW on the newer ones is worth noting. If you built your compliance review around the Nemotron Open Model License, the Ultra, Lightning and Bolt releases need a fresh read.

Architecture in plain terms

All the Nemotron 3 cards describe a hybrid design. Most layers alternate between Mamba-2 blocks and mixture-of-experts blocks, with a few attention layers. Mamba layers handle long sequences with memory that grows more slowly than attention, which is why NVIDIA can offer 1M-token windows at these sizes. Ultra uses what NVIDIA calls LatentMoE, where tokens are projected into a smaller latent dimension for expert routing. The larger models add multi-token prediction layers, which speed up generation, and Ultra's card says it was pre-trained with an NVFP4 recipe on about 20 trillion tokens.

Mixture of experts means only part of the model runs for each token, 55B of 550B in Ultra. That makes serving cheaper per token than a dense model of the same size. It does not make the model smaller in memory: all 550B parameters must be resident. That is why the hardware column is large.

Reasoning is switchable. The cards describe turning thinking on or off in the chat template, and the Lightning card mentions speculative decoding methods released alongside the model.

Context and modalities

Ultra, Super and Lightning each list up to 1M tokens. Bolt lists up to 256K and accepts text, images and video, outputting text. Ultra's card lists English, French, Spanish, Italian, German, Japanese, Hindi, Korean, Brazilian Portuguese and Chinese. Super lists a slightly shorter set of seven languages.

A caution about the context numbers: the Lightning card says it supports up to 1M tokens and uses 256K for a single-H100 deployment. That is a general pattern. The window is a design maximum. What you can use depends on how much memory is left after the weights.

Vendor-reported benchmarks

NVIDIA prints comparison tables on the cards. These are NVIDIA's own runs, with competitor figures chosen and reported by NVIDIA. We copy a few coding and long-context rows and label them accordingly.

Nemotron 3 Ultra card, vendor-reported:

Nemotron 3.5 Lightning card, vendor-reported:

Sources: the Hugging Face model cards for Nemotron 3 Ultra and Nemotron 3.5 Lightning, vendor-reported.

Read these tables honestly. Ultra is not the top scorer on coding in NVIDIA's own comparison: its 70.7 on SWE-Bench Verified sits below every other model in the row. Where it is strong is long context, with the best RULER at 1M in that table. Lightning is the same story at small size: in NVIDIA's table, Qwen's 35B-A3B model beats it on SWE-bench Verified and Terminal-Bench. So if the question is "which open model codes best," the Nemotron cards themselves say look elsewhere, and our coding shortlist does. The case for Nemotron is licence clarity, NVIDIA hardware efficiency and long context.

Running it

Bolt is the one that matters for a developer's desk. NVIDIA describes it as a native 2-bit quantised vision-language model designed to run everyday agent tasks locally alongside other applications, with speculative decoding through a separate small drafter. It trades precision for fit, and quality at 2 bits will differ from the full-precision models on the other cards. We have not seen NVIDIA or an independent test compare Bolt with its full-precision sibling..

The Lightning card says its BF16 release is meant mainly as a starting point for post-training, distillation and building your own quantised variants. For ordinary serving, pick the NVFP4 build. For tool choice, see llama.cpp versus Ollama and Ollama versus vLLM. For renting GPUs, see Lambda, Runpod and Modal.

Picking by hardware budget

Open-model choices are mostly decided by what you can afford to keep running. This list ranks the family by the memory you need, with the hardware each card names.

1. A single 24 GB GPU. Only Bolt, in its 2-bit GGUF form, is built for this. Expect to give up some precision.
2. A single H100. Lightning at 256K context, per its card. The NVFP4 build is the fast one.
3. One 8-GPU H100 node. Super at 120B. This is the sweet spot NVIDIA names for steady agent workloads with long inputs.
4. A Blackwell or H200 node. Ultra on 8 GPUs, or on 16 H100 across two nodes.

A rule of thumb for budgeting: GPU rental is billed by the hour, so a node that sits idle is the expensive failure. Rent for the length of an experiment, measure tokens per second and passes on your own tasks, then decide whether to buy. Our Runpod versus Lambda comparison shows how the hourly rates differ between hosts.

Customising a Nemotron model

NVIDIA publishes the pieces for customisation. The cards link pre-training datasets and a post-training dataset on Hugging Face, and the Lightning card links reproducibility recipes in NVIDIA's NeMo Gym repository. Both licences allow derivative works. Under the Nemotron Open Model License, if you distribute a derivative you must pass on the licence and include the attribution line. OpenMDW 1.1 has no separate derivative clause: if you distribute any part of the model materials you must include a copy of the licence and the applicable copyright and origin notices.

Start with the BF16 reference release if you plan supervised fine-tuning, reinforcement learning or distillation, and keep your own held-out test set. A tuned model can look better on your training tasks and worse on general ones, and you want to know which.

Where it is served

NVIDIA runs a hosted catalogue at build.nvidia.com, and its blog and docs name partner clouds. We did not extract per-token prices for any Nemotron model. The build.nvidia.com catalogue lists Nemotron 3 models such as nemotron-3-ultra-550b-a55b and several Nemotron 3.5 models, and the page states no per-token prices. Because the licences allow it, open-model hosts can serve the weights. See Hugging Face Inference, Replicate, Cloudflare Workers AI and AWS Bedrock, though we did not check which of them carry a Nemotron 3 model. For the general picture, see AI hosting.

Coding agents that can use it

Nemotron models can sit behind any agent that accepts an OpenAI-style endpoint, which you get from vLLM or SGLang if you self-host. The Ultra card shows benchmark runs on agent-style tasks, and the Lightning card refers to harness-level coding-agent results in a chart. The Ultra card says its evaluations used three main harnesses, Nemo Gym, Nemo Skills and Harbor, and the Lightning card says SWE-Bench and Terminal-Bench used NeMo Evaluator natively without naming a coding-agent harness. For tools with a custom provider setting, look at OpenCode, Cline and Aider. For Bolt specifically, the llama.cpp server is an OpenAI-compatible endpoint you can point those tools at. We have not tested Bolt's tool-calling inside those agents. Compare tools on choose an AI coding agent.

Privacy

Self-hosting a Nemotron model keeps all data on your own hardware, which is the main reason enterprises pick open weights. NVIDIA's cards describe the training data sources, listing public and synthetic datasets, and note a global deployment geography. If you use a hosted service instead, its own terms apply, and we did not read NVIDIA's hosted catalogue terms.

Limits and gotchas

• Weights-only size. Ultra needs 8 to 16 data-centre GPUs.
• Licence changed between releases. Super and Nano carry NVIDIA's licence, newer models OpenMDW.
• Coding scores trail rivals in NVIDIA's own tables.
• Bolt is 2-bit. It fits 24 GB by giving up precision.
• BF16 is for customisation. Use NVFP4 for fast serving of Lightning.
• Context is a maximum. Memory sets the usable length.
• Newer than most tooling. Hybrid Mamba models need up-to-date serving stacks. The Nano 30B card installs vllm>=0.12.0, the Nano 4B card asks for vllm>=0.15.1, and the Lightning and Ultra cards name tested containers, vLLM 0.27.1 and 0.22.0, rather than minimums.

Our take

Nemotron is the family to look at when your constraints are NVIDIA hardware, a clean licence and long context rather than the top coding score. Super on a single 8-GPU node is a reasonable private deployment for agent workloads with long inputs. Bolt is a curiosity worth testing on a 24 GB card if you want a local agent model with vision. Ultra is for organisations that already own a rack and want a very large open model.

If you want the strongest open coding model, the cards themselves point to others, such as Qwen, DeepSeek and Kimi. We would test Nemotron against them on your own repository before you buy GPUs.

Who should pick what

Related pages

• Open source LLM for licence comparisons
• Best open source LLM for coding for the shortlist
• Qwen models and Gemma models for neighbours in NVIDIA's own tables
• Cloud versus self-hosted AI for the cost maths