NVIDIA Nemotron 3.5, RTX Spark and Hugging Face: NVIDIA’s Bid to Put Workplace AI on the Desk

The usual way to use artificial intelligence (AI) at work is through a service such as ChatGPT or Claude: the text a professional types travels to the provider’s data centres, the model runs there, and the firm pays a subscription or a fee per use. NVIDIA Nemotron 3.5, a family of open AI models released on 15 October, includes one built to run on the laptop itself instead. Laptops with NVIDIA’s new RTX Spark chip went on sale the next day from $2,599, and NVIDIA says that model runs on supported configurations.

An open model is one whose trained files anyone can download, run and adapt, and NVIDIA’s are free to download and cleared for commercial use. Running one on a laptop means a document can be analysed without leaving the machine, and most of the cost of using AI becomes the price of the computer instead of a running bill. NVIDIA also agreed in September to buy Hugging Face, the platform where developers publish and download open models. The same company sells the chip, gives away the model and has agreed to buy the place where models are found.

Local AI becomes one more option for a law firm, a consultancy, a bank’s finance team or a fintech start-up, alongside the cloud services they already use. It is one company’s launch, and how many firms take it up remains to be seen. NVIDIA’s three moves show where it wants the decision about AI at work to be made: at the point a company chooses its computers.

What NVIDIA released on 15 and 16 October

The largest of the new models, Nemotron 3.5 Super VL, has 120 billion parameters, a rough measure of a model’s size, and reads images as well as text, which suits documents, charts and tables. NVIDIA published its weights, the trained files that make a model work, together with the data and instructions used to train it, according to its Nemotron 3.5 announcement. Sharing the training data as well goes further than publishing the weights alone. The company says Super VL is comparable with its own largest reasoning model at about a fifth of the size, and the one independent ranking published at launch, from Artificial Analysis, does not test the model’s image reading.

The laptop model, Nemotron 3.5 Bolt 74B-A9B, is tuned for agents, software that carries out a series of tasks on its own, such as researching the web, drafting text or fixing code. Its Hugging Face model card clears it for commercial use under the OpenMDW licence, a licence for AI models published by the Linux Foundation and submitted in August for approval by the Open Source Initiative. NVIDIA says the models are built for systems running its own graphics processing units (GPUs). A third, smaller model for voice assistants is due on 23 October.

RTX Spark personal computers (PCs) come from Dell, HP, Lenovo, Microsoft, ASUS and MSI, with up to 128 gigabytes of memory. NVIDIA’s RTX Spark launch release says AI agents can run on the device and that developers can keep the code and context they process locally. Microsoft calls Windows a platform for hybrid intelligence, meaning some AI tasks run on the device and others connect to the cloud.

Why NVIDIA gives a model away

NVIDIA earns its money from chips, and several of its recent deals support demand for them. It agreed in August to cover lease and power payments at an Ohio data centre leased to OpenAI if OpenAI fails to pay them, a deal we covered in NVIDIA’s Ohio backstop. A free model built for NVIDIA hardware serves the same purpose, as a reason to choose that hardware.

The pressure comes from NVIDIA’s largest customers. Microsoft said on 12 October that its own chip will become the primary system for AI inside its products from 2027, as set out in our piece on Microsoft’s Maia 300. Every cloud provider that designs its own silicon can buy fewer of NVIDIA’s chips, and a laptop on a professional’s desk is a new place to sell them.

Hugging Face sits between the cloud and the desk. NVIDIA’s Hugging Face acquisition, announced on 3 September at $12.93 billion, covers a platform used by over 200,000 companies, and by NVIDIA’s account it is already the largest contributor there, with more than 500 models. NVIDIA has pledged that the platform will stay open to every model builder and every type of hardware. NVIDIA’s filing on the deal expects it to complete in the first six months of 2027, once regulators in the United States and the European Union have approved it.

The same strategy reaches the developers who build on those models. NVIDIA’s Developer Program counts 200,000 enrolled developers in the United Kingdom alone, according to the one-year progress report on its UK sovereign AI commitments.

What changes in the AI bill

Cloud AI providers charge by the token, the unit of text a model reads or writes. OpenAI’s GPT-6.1 Sol, which sits between its GPT-6 Astra and GPT-6 Luna models on price, costs $2 per million tokens sent and $10 per million returned, according to the GPT-6.1 Sol launch page.

The final bill depends on how many tokens each task uses. Our analysis of GPT-5.6 pricing showed OpenAI keeping the rate of an earlier model unchanged and pinning the promised savings on the model using fewer tokens.

Anthropic took a different route, cutting the price it charges when its models reread material they have already processed, the change behind the AI cost finance pays. Both approaches keep the bill metered, so it grows with every task a team hands to the model.

A model running on a company’s own laptop works differently: the firm pays once for the machine, and each extra task adds little beyond electricity. NVIDIA has not said which RTX Spark configurations count as supported for Bolt. Tech Insider lists the cheapest machine in the range, Microsoft’s Surface Laptop Ultra at $2,599.99, with 24 gigabytes of memory, the same size as the model, so a laptop that runs it comfortably is likely to cost more.

The 128-gigabyte Surface Laptop Ultra lists at $5,899.99. That price buys, by our count, about 590 million tokens of output from GPT-6.1 Sol, and the $2,599 entry price buys about 260 million. The comparison leaves out the quality gap between models, setup and support, and it shows how differently the two ways of paying for AI behave once usage grows.

What it means for confidential work

Lawyers, auditors and bankers handle documents that clients expect to stay inside the firm. A model that processes a contract on the laptop where it is stored removes one step from the chain of who sees the data, and NVIDIA says code and context can remain on the device. The hybrid design Microsoft describes still sends some work to the cloud, so the gain depends on how each firm configures it.

Finance has its own open models to choose from. Ant Group released an open financial model built for research tasks such as valuation and report drafting. NVIDIA’s release adds an open finance dataset, and Ramp, the spend management firm, is adapting Super VL to read financial documents. Scale AI has trained versions of it for an EY agent that supports due diligence on mergers and acquisitions.

Where NVIDIA’s models meet the competition

NVIDIA enters a crowded field of open models. Mistral, the French AI developer, plans to release the weights of Mistral Large 4 by the end of October, and Chinese developers including DeepSeek and Alibaba’s Qwen team also publish open weights.

Super VL is also available through OpenRouter, a service that lets developers reach over 400 models from one account and picks between them on cost and speed. Stripe agreed in August to buy the service for more than $7 billion, according to Bloomberg, though the companies did not disclose terms, as covered in our piece on the OpenRouter acquisition. A router decides which model each request reaches, and so which developer gets paid for it.

The concentration question follows NVIDIA into every layer. Our coverage of Italy’s network failure examined the risk of building on too few providers, in networks and in AI. NVIDIA’s neutrality pledges for Hugging Face and the merger reviews in Washington and Brussels are where that risk will be tested.

NVIDIA is offering firms a way to bring AI into confidential work through a purchase order for a computer, and once the Hugging Face deal closes, expected in the first six months of 2027, the chip inside that computer, the model it runs and the shelf where the model is found will all lead back to the same company.

Editor’s note

Every piece goes through careful verification, but mistakes can happen. Readers who spot an error or have additional information can write to rosalia@thebrightminded.com.