Nvidia Nemotron 3.5 peut faire tourner un agent IA sur un PC équipé d’un GPU NVIDIA disposant d’environ 24 Go de mémoire vidéo, grâce au modèle Bolt 74B-A9B. Ce fonctionnement local convient à la recherche, à l’analyse documentaire et à certains assistants métiers. Il exige toutefois une configuration technique, des contrôles humains et un budget matériel à comparer au coût d’un service cloud.
What is Nvidia Nemotron 3.5?
Nvidia Nemotron 3.5 is a family of artificial intelligence models capable of reasoning, analyzing documents, and using tools. Launched on October 15, 2026, it notably includes Super VL 120B-A12B for demanding multimodal workloads and Bolt 74B-A9B for local execution with approximately 24 GB of memory.
An AI agent is software that combines a model with tools to complete several steps: viewing files, conducting a search, calling an application, or producing a structured result. Unlike a simple chatbot, an agent can chain actions according to a goal and partially monitor their outcome.
Nemotron 3.5 Super VL is a vision-language model, meaning that it processes text, images, and video. According to NVIDIA in 2026, it has 120 billion parameters in total, of which 12 billion are activated for each processing operation, and supports a context window of up to one million tokens. A token corroughly corresponds to a word fragment.
This theorretical maximum does not guarantee that a workstation can process one million tokens under good conditions. The official documentation specifies that the practically usable length and the number of simultaneous requests must be reduced according to the available memory.
Bolt 74B-A9B is a compressed version derived from Super VL. Its 74 billion parameters, 9 billion of which are active, use native quantization of approximately 2 bits: the model’s internal numbers are stored with less precision to significantly reduce memory usage. This trade-off reduces the hardware footprint but may also affect the quality of some responses.
Can Nvidia Nemotron 3.5 really run on a PC?
Nvidia Nemotron 3.5 Bolt can fit within a 24 GB memory budget according to NVIDIA in 2026, making local execution possible on certain PCs equipped with a compatible RTX GPU. This does not mean that a standard corputer is sufficient: the software, system memory, storage, cooling, and agent tools still need to be configured.
A GeForce RTX 5090 has 32 GB of GDDR7 memory according to its official 2026 specifications. Its capacity therefore exceeds the memory budget announced for Bolt, but the memory remaining for the context, cache, and simultaneous processing remains limited. A 24 GB card leaves evoren less headroom.
The word “local” can sometimes be confusing. The model can run on your own equipment without sending every request to an inference provider, but the agent can still call the Internet, an API, or a search engine if its workflow calls for it. Its actual level of autonomy therefore depends on the complete architecture, not the model alone.
In the projects we undertake, we often see GPU memory become the sole purchasing criterion. This is a mistake: drivers, the operating system, inference libraries, and the ability to supervise the service also determine its stability. The independent test of a Nemotron 3.5 Lightning model on a ortable RTX 5070 Ti 12 GB corputer, published on GitHub in August 2026, also documents substantial configuration work and offloading to the processor.
Three variants illustrate very different levels of equipment:
| Model | Announced footprint | Hardware cited | Relevant use case |
|---|---|---|---|
| Super VL 120B-A12B | 120 billion parameters, 12 billion active | 1 DGX Spark, 1 H200, 2 H100s, or 1 GB200 in NVFP4 | Documents, images, video, code, and complex workflows |
| Bolt 74B-A9B | 74 billion parameters, 9 billion active, 24 GB budget | NVIDIA RTX GPU or DGX Spark | Research, writing, debugging, and local tool execution |
| Bolt E4B 4GB | Announced lightweight variant, with no detailed public specifications as of October 15, 2026 | GeForce RTX, Jetson, and RTX or DGX Spark | Conversational, voice, or embedded agents |
What uses are realistic for an SMB?
For an SMB, Nvidia Nemotron 3.5 is mainly useful for document analysis, assisted research, image or video understanding, and the execution of internal tools. Bolt reduces the required infrastructure, while Super VL targets more demanding workloads. Neither model eliminates the need to verify business results.
One initial use case is querying a collection of contracts, procedures, or technical files stored on controlled infrastructure. The model can search for a clause, summarize multiple documents, or prepare a table. However, each response must be linked to its source documents, and the agent must be prevented from accessing directories unrelated to its mission.
The second use case concerns development assistance: code explanation, error detection, and preparation of modifications. A trial published by MindStudio on October 17, 2026, showed good results on a visual reasoning task but failures in a cor software correction and HTML generation. The sample is too small to measure overall reliability, but it confirms that human validation remains necessary in production.
Autonomous translation or direct content publication are also risky. The same independent trial identified weaknesses in multilingual sorties. Frankly, an agent that can modify a website or send a document to a client should not have these permissions without an approval step.
Before deployment, use cases can be classified according to their level of impact:
- low risk: internal research, drafting, summarization, and classification without deleting the originals;
- moderate risk: code generation or contract analysis reviewed by a qualified person;
- high risk: payments, data deletion, publication, or messages automatically sent to a client;
- risk to prohibit without a dedicated system: unrestricted access to the network, passwords, or data belonging to multiple clients.
To compare this family with another European offering, the analysis of the Mistral Large 4 model provides another perspective on the trade-offs between capabilities, hosting, and operation. The choice should not be based solely on a benchmark ranking.
What budget should you plan for a local AI agent?
The budget for Nvidia Nemotron 3.5 is not limited to downloading the model. In 2026, a 32 GB GeForce RTX 5090 has a U.S. list price of $1,999 to $1,999.99, while an NVIDIA DGX Spark is listed at $4,699 in the United States. French prices, integration, and maintenance are additional.
The DGX Spark includes 128 GB of unified memory, a 4 TB SSD, and a GB10 chip rated for a 140 W thermal design power according to NVIDIA in 2026. Its U.S. retail price rose from $3,999 to $4,699 in February 2026. Taxes, availability, and reseller pricing make any direct conversion to euros unreliable.
The price of an RTX 5090 is no more the total cost of a workstation. You also need a compatible corputer, an appropriate power supply, storage, installation time, monitoring, and updates. A failure also brings the service to a standstill if no second device or recovery system is available.
At this budget, it is better to start with a prototype limited to a corpus and a reversible action rather than immediately building a general-purpose agent. Two to four weeks of internal testing provide more useful information than a hardware purchase based solely on a vendor demonstration. This timeframe is a project-planning guideline, not a measurement published by NVIDIA.
The cloud is often more practical for occasional use, testing, or highly variable workloads. Local deployment becomes attractive loren confidentiality, latency, recurring volume, or infrastructure control outweigh the purchase and operating costs. For example, Plotly reported in October 2026 a cost of $0.46 for its hosted evaluation of 43 questions with Super VL; this result measures neither a local PC nor annual operating costs.
A local architecture also requires a continuity plan. The principles of a disaster recovery plan remain useful for determining how to restore the model, its document indexes, and its access rules after an outage.
What performance and risks should be anticipated?
The published performance figures for Nvidia Nemotron 3.5 still come primarily from NVIDIA as of October 2026. According to the provider, Bolt reaches 83.99 tokens per second on AIPerf SPEED-Bench-Coding with LiLiCorr and RTX Spark hardware. This figure is not an independent benchmark of a Windows or Linux PC equipped with 24 or 32 GB.
As of October 21, 2026, no extensive independent study has been identified on Bolt’s sustained speed, power consumption, total cost, or maintenance on ordinary PCs. Community tests exist on RTX 3090 and RTX 5090 systems, but their configurations are not standardized. The speed claim should therefore be validated using your own documents and tools.
Local security is not limited to keeping files on your premises. A malicious document may attempt to influence the agent, a misconfigured connector may expose a database, and an overly permissive action may overwrite information. From an agency perspective, the best practice is to separate read and write access, then log every sensitive action.
NVIDIA itself asks organizations to maintain equivalent safeguards and assess the risks of misuse, confidentiality, and security for each application. The priority controls are authentication, least-privilege access, encryption, logs, and human validation. The free cybersecurity control tools can complement an initial audit without replacing an architecture review.
Visit RGPD s’applique dès qu’un agent traite des données personnelles, même si le calcul reste local. Il faut définir la finalité, limiter les données, fixer une durée de conservation et documenter les accès. Pour mieux mesurer le risque lié aux fournisseurs et sous-traitants, le retour d’expérience sur la securing data entrusted to a service provider provides a useful framework.
Finally, according to the 2026 documentation, the OpenMDW-1.1 license authorizes commercial and noncommercial use. A permissive license does not transfer deployment responsibility to the publisher. NVIDIA also states that some information relating to the training data is not disclosed and recommends an assessment specific to each application.
Defining the scope of a local agent in advance helps avoid buying a machine that is too limited or granting the model too many permissions. An outside perspective can be especially helpful when comparing a local workstation, a DGX Spark, and a hosted infrastructure based on the same business scenario.
FAQ about Nvidia Nemotron 3.5
Does Nvidia Nemotron 3.5 work without an Internet connection?
Nvidia Nemotron 3.5 Bolt can run the model locally without sending requests to a cloud inference service. A connection is still required if the agent uses web search, an external API, or hosted tools.
What graphics card is required for Nemotron 3.5 Bolt?
Nemotron 3.5 Bolt is designed to fit within a memory budget of approximately 24 GB, according to NVIDIA in 2026. However, a 24 GB card represents a tight limit, while a 32 GB GeForce RTX 5090 provides more headroom for context and processing.
Can Nemotron 3.5 analyze PDFs and images?
Nemotron 3.5 Super VL prend en charge l’analyse documentaire, les images et la vidéo selon NVIDIA en 2026. L’exploitation de PDF exige une chaîne d’importation, des droits d’accès et un mécanisme permettant de rattacher les réponses aux documents sources.
Can Nemotron 3.5 be used in a company?
Nemotron 3.5 is offered under the OpenMDW-1.1 license for commercial and noncommercial uses in 2026. The company remains responsible for GDPR compliance, security, human validation, and the actions entrusted to the agent.