NVIDIA Nemotron 3.5 Explained: The Free Open AI Models Built for Agents (2026 Guide)

NVIDIA Nemotron 3.5 Explained: The Free Open AI Models Built for Agents (2026 Guide)

On October 15, 2026 — just three days ago — NVIDIA quietly did something that has the whole AI world talking: it released NVIDIA Nemotron 3.5, a brand-new family of three open-weight AI models built from the ground up for agents. The weights are free to download, commercial use is allowed, and one of the models squeezes 74 billion parameters into a single 24GB gaming PC. If you build AI agents, run local LLMs, or just want to understand where open AI is headed next, this is the launch you need to know about.

In this guide, we break down everything NVIDIA Nemotron 3.5 includes: the three models, their real specs, what the benchmarks actually say (with honest caveats), the OpenMDW-1.1 license in plain English, how they stack up against Meta, Google, and Chinese open models, and how you can try them yourself today — without paying a cent for the weights.

NVIDIA Nemotron 3.5 open AI models running on RTX GPU hardware

Table of Contents

What Is NVIDIA Nemotron 3.5?

NVIDIA Nemotron 3.5 is a family of three open AI models that NVIDIA announced on October 15, 2026, designed specifically for the agent era — AI systems that do not just answer questions but take actions: browsing the web, calling tools, reading documents, running code, and carrying multi-step workflows forward on their own. The model weights were posted on Hugging Face the same day the announcement went live on NVIDIA’s blog.

Think of it this way: most famous LLMs are built to be good conversationalists. NVIDIA Nemotron 3.5 is built to be a good worker. NVIDIA’s whole pitch is that these models are optimized for instruction following, tool use, long business workflows, and visual understanding of documents, charts, and screenshots — the unglamorous skills that actually make agents useful in a real company.

Open-Weight vs. Open Source: What “Open” Really Means Here

You will see people call NVIDIA Nemotron 3.5 “open source.” That is almost right, but precision matters. What NVIDIA released are open weights: the trained model parameters are free to download, use, modify, and deploy — including for commercial products — under a license called OpenMDW-1.1. That is the part most developers care about.

Strictly speaking, “open source” in the OSI sense also implies open training data and full build transparency. NVIDIA goes further than most vendors here by shipping open datasets and training recipes alongside the weights, so teams can post-train the models on their own private data. But it is still more accurate to say open-weight than open source. We will unpack the license terms in Section 4.

Where Nemotron 3.5 Fits in the Nemotron Family

One quick clarification, because this gets confusing: NVIDIA Nemotron 3.5 has nothing to do with the 2023 “Nemotron-3” enterprise models. Totally different generation. In the current lineup, the new models join two siblings — Nemotron Lightning, tuned for maximum throughput on high-volume workloads, and Nemotron Ultra, NVIDIA’s heavyweight for the hardest reasoning tasks. Nemotron 3.5 sits in the middle of the action: the agentic tier, covering everything from datacenter servers down to edge devices and local PCs.

NVIDIA Nemotron 3.5 models deployed in a data center for always-on agents

Meet the Three NVIDIA Nemotron 3.5 Models

The NVIDIA Nemotron 3.5 family is really three different products aimed at three different machines: a flagship for the datacenter, a desktop powerhouse, and an edge lightweight. Here is each one in detail.

Nemotron 3.5 Super VL — The 120B Agent Flagship

Nemotron 3.5 Super VL is the headliner: a 120-billion-parameter mixture-of-experts (MoE) model with 12 billion active parameters per token, a 1-million-token context window, and true multimodal input. It reads text, images, documents, and video, then reasons and calls tools before producing text output. It is available now.

NVIDIA built Super VL — in collaboration with what it calls the Nemotron Coalition — for specialized, always-on agents: the kind that sit inside a company and quietly grind through instruction following, tool use, multi-step business workflows, and visual understanding all day. Think finance agents reconciling documents, support agents reading screenshots, or research agents working through long PDF reports. NVIDIA is already pointing at early adopters: Scale AI is using it for finance agents, AfterQuery for credit-agreement analysis, and TrainLoop for dermatology prescription drafting. Platform partners including Atlassian, CrowdStrike, Salesforce, and Perplexity are integrating the models into their own agent products.

Because it is a sparse MoE design, only 12B of the 120B parameters fire on any given token — so per-token compute looks more like a 12B model — but the full 120B still has to live in memory. NVIDIA names the hardware it expects: a single DGX Spark, one H200, two H100s, or one GB200. That is serious datacenter territory, not a gaming PC.

But the payoff NVIDIA claims is serious too: Super VL matches the accuracy of NVIDIA’s largest reasoning model, Nemotron 3 Ultra, at roughly one-fifth the parameter count, and delivers up to 5x higher agentic throughput on a single GPU compared with other open models in its class. Those are NVIDIA’s claims, tied to the Artificial Analysis leaderboard — we will put them under the microscope in Section 3.

One of the most interesting parts of the release: Super VL ships with open weights, open datasets, and training recipes. That means an enterprise team can take the base model and post-train it on their own private documents and workflows — the model becomes a specialist in their business, not just a generalist. For companies that cannot send sensitive data to a cloud API, this is the whole ballgame.

Nemotron 3.5 Bolt 74B-A9B 24GB — The Desktop Beast

If Super VL is the datacenter flagship, Nemotron 3.5 Bolt 74B-A9B 24GB is the model most individual developers will actually touch. It packs 74 billion total parameters with 9 billion active per token — and, crucially, it uses native 2-bit quantization so the entire model fits in 24GB of memory. That is exactly the VRAM on a GeForce RTX 5090. Also available now, it is the first 74B-class open model that NVIDIA says runs comfortably on a high-end consumer PC, a DGX Spark, or RTX Spark laptops.

Bolt 74B is built for local agentic work: web research, writing, and code debugging without touching the cloud. Everything stays on your machine — no API keys, no per-token bills, no sending client documents to a third party. With a 256K-token context window, it can hold very long documents or codebases in memory at once.

The numbers NVIDIA published are eye-catching: 50.98 on Terminal-Bench 2.1 (an agentic coding benchmark) and 86.43 on GPQA Diamond (a hard science-reasoning test). Again: these are vendor-reported figures, not independent lab results. Still, they suggest a model that punches well above what you would expect from something that runs on a desktop.

Practically, Bolt 74B meets developers where they already live. NVIDIA lists compatibility with llama.cpp, Ollama, LM Studio, Unsloth, and vLLM — the standard local-LLM toolkit. If you already run open models locally, you know these names; the new part is getting this much capability in 24GB.

Developer running NVIDIA Nemotron 3.5 Bolt locally on an RTX workstation

Nemotron 3.5 Bolt E4B 4GB — The Edge Lightweight

The third member is Nemotron 3.5 Bolt E4B 4GB, and it is not out yet — NVIDIA says it arrives October 23, 2026, eight days after the other two. This is the small language model of the family, sized for fast conversational and physical agents: local voice assistants, interactive characters, and embedded interfaces running on GeForce RTX cards, Jetson modules, and RTX/DGX Spark systems.

The target here is intelligence-per-gigabyte: snappy, on-device agents that do not need a datacenter or even a beefy desktop. Robotics, kiosks, in-car assistants, smart-camera systems — anywhere an agent needs to live inside the device. We will know more when the weights drop on the 23rd.

NVIDIA Nemotron 3.5 Models Compared

Model Parameters (active) Context Hardware Best for Availability
Super VL 120B (12B active) 1M tokens DGX Spark / H200 / 2× H100 / GB200 Always-on enterprise agents, document + video understanding Now
Bolt 74B-A9B 24GB 74B (9B active, 2-bit) 256K tokens RTX 5090 PC, DGX Spark, RTX Spark laptops Local coding, research, writing agents Now
Bolt E4B 4GB Not disclosed Not disclosed GeForce RTX, Jetson, RTX/DGX Spark On-device conversational + physical agents Oct 23, 2026

Key Specs of NVIDIA Nemotron 3.5

NVIDIA Nemotron 3.5: The Numbers That Matter

Here is the spec sheet, stripped of marketing gloss:

  • Architecture: mixture-of-experts (MoE) across the announced models — only a slice of parameters activates per token (12B of 120B in Super VL, 9B of 74B in Bolt), which keeps per-token compute low even though total parameter counts sound huge.
  • Context windows: 1 million tokens for Super VL; 256K tokens for Bolt 74B. In practical terms, 1M tokens is roughly an entire codebase, a shelf of contracts, or hours of video transcripts in one prompt.
  • Multimodal input: Super VL accepts text, images, documents, and video as input (text output). Bolt 74B is aimed at text-centric agentic work like research, writing, and debugging.
  • Quantization: Bolt 74B uses native 2-bit quantization — most parameters compressed to about 2 bits each — which is how 74B parameters fit into 24GB of memory. Compression always involves tradeoffs, but native 2-bit (trained for it, not bolted on afterward) holds up better than post-training quantization.
  • Tool use + reasoning: both available-now models are built for agentic loops — reasoning, then calling tools, then reasoning again — rather than one-shot answers.
  • License: OpenMDW-1.1, commercial use allowed (details in Section 4).
  • Availability: weights on Hugging Face, plus OpenRouter and ModelScope; also appearing in hosted catalogs.

What the Benchmarks Actually Say

Now the part that needs a clear head. Every benchmark number in this section is NVIDIA’s own claim or vendor-reported data — not an independent audit. With that caveat on the table:

  • Super VL: NVIDIA says it matches its largest reasoning model (Nemotron 3 Ultra) at roughly one-fifth the parameters, and delivers up to 5x higher agentic throughput on a single GPU versus other open models in its class, citing the Artificial Analysis leaderboard. If those numbers hold up in independent testing, Super VL would be the efficiency leader for always-on agents — more completed tasks per GPU-hour, which is the metric enterprises actually pay for.
  • Bolt 74B: NVIDIA reports 50.98 on Terminal-Bench 2.1 and 86.43 on GPQA Diamond. Terminal-Bench measures real agentic terminal work (running commands, fixing code, navigating a machine); GPQA Diamond is a brutally hard science benchmark. Scores in this range from a 24GB desktop model would have been unthinkable two years ago.

What is missing? Independent, head-to-head testing against the models NVIDIA is implicitly competing with — and the usual fine print about test conditions, sampling, and cherry-picked prompts. Treat the numbers as a strong signal of intent (NVIDIA is optimizing hard for agentic throughput, not just chat quality) and wait for the community benchmarks on Hugging Face leaderboards before treating them as settled fact. The good news: because the weights are open, anyone can verify them. That is the entire point of open weights.

The License: What OpenMDW-1.1 Lets You Do

The NVIDIA Nemotron 3.5 license is called OpenMDW-1.1, and the short version is encouraging: the weights are free to download, modify, and deploy, and commercial use is allowed. You can build a product on these models, sell it, and run it for clients without paying NVIDIA a license fee.

That said, “open weights with a custom license” is not the same as public-domain software. As with every model license, you should read the actual OpenMDW-1.1 terms before shipping a product — particularly the sections on acceptable use, redistribution, and attribution. Licenses like this typically allow commercial deployment but reserve NVIDIA’s trademarks and set conditions around how the models can be offered as a service. The weights are free; the legal homework is on you. If your company has a legal team, have them spend an hour on the license text before you commit a roadmap to it.

The practical takeaway: for indie developers, researchers, and startups, OpenMDW-1.1 is permissive enough to build on with confidence. For large enterprises, it is permissive enough to evaluate seriously — which is exactly why NVIDIA also ships the open datasets and training recipes. A bank or hospital can take Super VL, post-train it on internal documents that never leave the building, and deploy agents on their own hardware. No API, no data leaving the premises, no per-token meter running.

NVIDIA Nemotron 3.5 vs the Competition

NVIDIA did not release these models into a vacuum. The open-model space in late 2026 is brutally competitive, and each rival has a different strength. Here is an honest map.

vs. Meta Muse Spark Open

Meta’s open releases (including the Muse Spark open line we covered in our Meta Muse Spark Open guide) compete on ecosystem and community fine-tunes — nobody has a bigger open-model community than Meta’s Llama lineage. NVIDIA’s counter is specialization: Nemotron 3.5 is tuned specifically for agents (tool use, long workflows, document vision) rather than general chat, and it is optimized to run best on NVIDIA hardware, which most serious deployments already use. If your workload is agentic and your GPUs are green, Nemotron 3.5 is arguably the more purpose-built choice; if you want the largest library of community fine-tunes and adapters, Meta’s ecosystem still leads.

vs. Elorian Glimpse

Elorian’s Glimpse (see our Elorian Glimpse breakdown) pushed the frontier on efficient open models earlier this year. NVIDIA’s answer with Super VL is a direct efficiency play of its own — Ultra-class accuracy at a fifth of the parameters, 5x the agentic throughput per GPU. On paper that is a stronger efficiency story than anything in the Glimpse class. But Elorian’s models earned trust through months of independent benchmarking; Nemotron 3.5’s numbers are three days old and vendor-reported. Revisit this comparison in a month, when the community leaderboards have spoken.

vs. Qwen, GLM, and the Chinese Open Models

Here is where honesty matters most: on several coding benchmarks, Chinese open models in the Qwen and GLM families still post higher scores than what NVIDIA has published for Nemotron 3.5 so far. If your primary job is raw code generation quality and you do not care about NVIDIA’s agentic tuning, those models remain extremely strong — and often smaller and cheaper to run. NVIDIA’s bet is that agentic throughput — how many real multi-step tasks a model completes per GPU-hour — matters more than raw benchmark scores for the agent era. It is a reasonable bet, but it is still a bet, and it is unverified by third parties as of this writing.

vs. Closed Models (GPT, Claude, Gemini)

Closed frontier models still lead on the hardest reasoning tasks, and Google’s latest releases (covered in our Gemini 4 Argon guide) show how fast the closed side is moving. But closed models come with API bills, data-leaving-your-servers, and zero ability to fine-tune the weights. Nemotron 3.5 is playing a different game: good enough intelligence that you fully own. For many businesses, owning the model beats renting a slightly smarter one — especially when the owned one runs on hardware they already bought.

For broader context on how the open-model landscape shifted this year, independent coverage of the launch is worth a read: Oossa’s launch report and Maccus’s technical breakdown both dig into the details with useful skepticism about vendor claims.

Who Should Actually Use These Models

Not everyone needs a 120B agent model. Here is who NVIDIA Nemotron 3.5 is actually for — and who can skip it.

Enterprise Teams Building Always-On Agents

If you run document-heavy workflows — finance, legal, insurance, healthcare admin — Super VL is the model to evaluate first. The combination of a 1M-token context, document and video understanding, tool use, and post-training recipes means you can build agents that read the messy real-world inputs your business actually produces: scanned contracts, charts, long email threads, recorded meetings. And because the weights are open, regulated industries can deploy on-premises without data ever leaving the building.

NVIDIA Nemotron 3.5 analyzing business documents and PDF reports for enterprise agents

Indie Developers and Power Users With RTX Cards

This is the most exciting audience. If you own a 24GB RTX 5090-class GPU — a serious but increasingly common enthusiast setup — Bolt 74B gives you a genuinely capable local agent for coding help, research, and writing, with zero API costs and total privacy. Students, freelancers, and small dev shops that could never afford datacenter inference suddenly get frontier-adjacent agentic capability on a desktop. The Ollama and LM Studio support means setup is measured in minutes, not days.

Edge and Robotics Builders

When Bolt E4B 4GB lands on October 23, it targets the builders putting agents inside things: Jetson-powered robots, retail kiosks, voice assistants, in-vehicle systems. If you work in embedded AI, mark the date — 4GB-class conversational agents that run fully on-device are exactly what the edge market has been waiting for.

NVIDIA Nemotron 3.5 Bolt E4B running on edge devices and robotics hardware

Who Can Skip It

If you just want a chatbot for casual questions, a smaller model (or a free hosted chatbot) is simpler. If you need the absolute best raw code generation and do not care about NVIDIA’s agentic tuning, the top Chinese open models may serve you better today. And if you have no GPU at all, start with the hosted options in the next section rather than buying hardware for one model.

How to Try NVIDIA Nemotron 3.5 Today

Two of the three NVIDIA Nemotron 3.5 models are downloadable right now. Here is how to get running, from easiest to most hands-on.

  1. Try it hosted (no hardware needed): the fastest path is a hosted endpoint. The models are listed on OpenRouter and ModelScope, alongside Hugging Face. Sign up, pick the Nemotron 3.5 model, and start chatting or calling the API within minutes. This is the best option if you have no GPU — you pay only for the tokens you use, and the weights themselves cost nothing.
  2. Download the weights from Hugging Face: search Hugging Face for the official NVIDIA Nemotron 3.5 repositories (posted October 15, 2026). The weights are free to download under the OpenMDW-1.1 license. Verify you are pulling from NVIDIA’s official organization page, not an unofficial re-upload.
  3. Run Bolt 74B with Ollama: if you have a 24GB GPU (RTX 5090-class), install Ollama, pull the Bolt 74B model, and run it from your terminal. Ollama handles the quantization and memory management for you — this is the simplest true-local setup.
  4. Run it with LM Studio (visual interface): prefer a point-and-click app? LM Studio on Windows or Mac lets you download Bolt 74B, load it into your GPU, and chat through a clean desktop UI — no command line required. Great for writers and researchers who just want a private local assistant.
  5. Advanced: llama.cpp, Unsloth, or vLLM: developers who want maximum control — custom quantization, fine-tuning, or high-throughput serving — can use llama.cpp for lean CPU/GPU inference, Unsloth for efficient fine-tuning, or vLLM for serving Super VL to a team. Super VL’s training recipes are published for teams that want to post-train on private data.
  6. Wait for Bolt E4B (October 23): if your target is edge devices, Jetson boards, or lightweight on-device agents, hold out for the 4GB model’s release on October 23, 2026, then follow the same Hugging Face download path.

A note on cost: the model weights are free — every path above starts at $0 for the software. Your costs are hardware (a 24GB-class GPU is a serious enthusiast purchase; datacenter GPUs like the H200 or GB200 are enterprise investments measured in the tens of thousands of dollars) or hosted per-token fees if you use OpenRouter-style endpoints. Never pay anyone for “access” to the weights themselves; they are public.

Honest Limitations

No launch is perfect, and three days is too early for final judgments. Here is what to keep in mind:

  • The benchmarks are vendor-reported. Every impressive number in this article traces back to NVIDIA. Independent testing takes weeks; until the community reproduces these results, treat them as promising claims, not facts.
  • Chinese open models still lead on some coding tests. If raw code-generation scores are your top priority, Qwen- and GLM-class models currently post stronger numbers on several coding benchmarks. NVIDIA is optimizing for agentic throughput, which is a different (and arguably more useful) metric — but do not assume Nemotron 3.5 wins every category.
  • Super VL needs serious hardware. A 120B model is not a laptop model. DGX Spark, H200, dual H100, or GB200-class hardware is datacenter or workstation territory with enterprise price tags. Most individuals should focus on Bolt 74B instead.
  • The license has conditions. OpenMDW-1.1 allows commercial use, but it is a custom license, not a public-domain grant. Read the terms — especially around redistribution and acceptable use — before building a business on it.
  • Bolt E4B is not out yet. Everything about the 4GB edge model is pre-release until October 23. Specs and real-world performance are still unknown.
  • 1M context is not free. Long contexts consume serious memory and compute at inference time. A 1M-token window is a powerful capability, but actually filling it is expensive — budget accordingly.

Verdict: Is NVIDIA Nemotron 3.5 Worth Your Time?

Yes — with the right expectations. NVIDIA Nemotron 3.5 is the most significant open-weight release for agents this year: a genuinely capable 120B multimodal agent model for the datacenter, a 74B model that fits on a desktop GPU, and a 4GB edge model arriving October 23. The weights are free, commercial use is allowed, the tooling support (Ollama, LM Studio, vLLM) is already there, and real companies are already building on it.

The honest caveat is timing: at three days old, the headline benchmark claims are still NVIDIA’s word, not the community’s verification. But that is precisely why open weights matter — anyone can test them. If you build agents, download Bolt 74B this weekend and put it through your workflows. That test is worth more than any launch-day benchmark.

Our take: for enterprise agent teams, evaluate Super VL now. For indie developers with a 24GB GPU, Bolt 74B is the most exciting local model release of 2026 so far. For everyone else, the hosted endpoints make it a zero-risk try. The agent era just got a major open foundation — and it is free.

FAQ

Is NVIDIA Nemotron 3.5 really free?

Yes — the model weights are free to download and use, including for commercial products, under the OpenMDW-1.1 license. You never pay for the weights themselves. Your costs are hardware (GPUs) or per-token fees if you use a hosted endpoint like OpenRouter.

Can I run NVIDIA Nemotron 3.5 on my home PC?

The Bolt 74B-A9B 24GB model is designed exactly for that: its native 2-bit quantization lets it fit in 24GB of VRAM, which matches a GeForce RTX 5090-class card. Super VL (120B) needs datacenter-class hardware like a DGX Spark, H200, or dual H100s. The 4GB Bolt E4B, arriving October 23, 2026, will run on even lighter devices including Jetson boards.

How is NVIDIA Nemotron 3.5 different from the 2023 Nemotron-3 models?

They share a name and nothing else. The 2023 Nemotron-3 models were an earlier enterprise lineup; Nemotron 3.5 is a new-from-scratch open-weight family built specifically for AI agents, joining the current Lightning (throughput) and Ultra (hard reasoning) tiers.

What is the difference between Super VL and Bolt?

Super VL (120B, 1M context) is the multimodal flagship for datacenter always-on agents — it reads text, images, documents, and video. Bolt 74B (24GB) is the desktop model for local coding, research, and writing agents. Bolt E4B (4GB, coming October 23) is the edge lightweight for on-device conversational and physical agents.

Is NVIDIA Nemotron 3.5 open source?

Strictly speaking, it is open-weight rather than open source: the trained weights are free to download, modify, and commercially deploy under the OpenMDW-1.1 license, and NVIDIA also publishes open datasets and training recipes. But it uses a custom license with conditions, so “open weights” is the accurate term.

How does NVIDIA Nemotron 3.5 compare to ChatGPT or Claude?

Different lane. Closed models like ChatGPT and Claude still lead on some raw reasoning tasks, but they are rented APIs: you pay per token, your data leaves your servers, and you cannot fine-tune the weights. Nemotron 3.5 offers strong agentic capability that you fully own and can run privately — for many businesses, ownership beats renting a slightly smarter model.

Leave a Comment

Your email address will not be published. Required fields are marked *