Nvidia has released three new open-weight AI models, according to TechCrunch. One is a large model for companies that run AI agents on their own systems. The other two are smaller Bolt models meant to run directly on a new class of AI PCs. The second part matters most for you. Nvidia's Kari Briski wrote that PC users "can delegate everyday projects … without sending those workloads to the cloud."
The background you need
AI agent. A chatbot answers questions. An agent takes action. You give it a goal, it breaks the goal into steps, and it uses tools to finish the job, such as searching the web, opening files or filling in forms. Tool use (or tool calling) is the term for a model triggering one of those actions. Always-on agents keep working in the background instead of waiting for you to type.
Parameters. These are the internal numbers a model learns during training. More parameters usually means a more capable model. It also means the model needs more computing power and memory. Size therefore decides where a model can run. A very large model needs a data center. A smaller one can fit on a powerful laptop.
Open-weight vs. closed. Weights are the trained parameters, the finished brain of the model. With a closed model, like the ones behind ChatGPT or Claude, you never get the weights. You send each request over the internet to the company's servers. With an open-weight model, the weights are published. An organization can download the model and run it on its own hardware.
Think of a closed model as a restaurant meal. You eat it there, on their terms. An open-weight model is the finished dish handed to you to reheat and season in your own kitchen. The analogy breaks in one place: you get the dish, not necessarily the full recipe. You may not see the training data or methods. Open-weight also tells you nothing about price or license terms. Those are separate questions.
Post-training. This means taking a model that is already trained and training it further on your own examples. The goal is to make it better at a narrow job, such as reading your company's invoices.
Edge computing and local AI. Edge means doing the computing close to where the data lives, on your PC or a device in the office, instead of in a remote cloud data center. When a model runs locally, your prompts and files are processed on your own machine.
Why a chipmaker builds models. Nvidia is best known for the chips (GPUs) that train and run much of today's AI. Models that run well on its hardware give customers a reason to buy that hardware. TechCrunch notes that Nvidia already has a whole Nemotron family of models. The family covers multimodal AI, reasoning, coding, math and enterprise automation. The labs these models compete with are also Nvidia's customers and investments.
TechCrunch also reports that those labs are now turning to other chip and system makers, such as AWS's Trainium and Cerebras, and are building their own chips. TechCrunch presents Nvidia's push into companies running their own open models against that backdrop.
What actually changed
Some of this was already true. Nvidia had released open-weight models before. Technical users could already run smaller models on their own computers. What is new, according to TechCrunch:
- Nemotron 3.5 Super VL. A 120-billion-parameter model built for enterprise agents. It has fewer parameters than Nvidia's Nemotron 3 Ultra. Nvidia says it delivers similar accuracy on instruction-following and tool-use tasks.
- Speed claims. Nvidia claims Super VL completes tasks faster than competing Anthropic and OpenAI models (Opus 5, GPT 6.1 Sol, Luna, Astra). That claim applies when Super VL is post-trained on specific agentic tasks. The tasks are document intelligence, financial accounting, scientific experimentation, medical insurance claims and security attack resilience. TechCrunch calls such claims standard for any major model launch and does not cite independent tests. Nvidia names companies it says use Super VL: Skydance, EY, Ramp, Crowdstrike, Perplexity and Salesforce.
-
Two Bolt models for local use. They are intended for new AI PCs running Nvidia's RTX Spark chip and for workstations running the DGX Spark.
- Nemotron 3.5 Bolt 74B-A9B 24GB is a 74-billion-parameter model for local agentic tasks. Nvidia says it balances speed with reasoning to handle multistep reasoning, tool calling and web browsing.
- The smallest model, Bolt E4B, is aimed at conversational agents.
- Availability. Nvidia says Super VL and Bolt 74B-A9B are available now, and Bolt E4B is coming later this month.
- The hardware. The AI PCs are being released now through partners like Microsoft and Dell. They are aimed at consumers, programmers and enterprise knowledge workers.
What it means for PMs, analysts and other knowledge workers
The local-processing promise outweighs the other details because it touches the rule that shapes most office AI use: what data you are allowed to send outside the company. Many organizations limit what employees can paste into cloud chatbots. If an agent runs on your laptop and the work stays there, the question IT has to answer changes. It does not answer the question for them.
Here are the concrete consequences.
Your laptop becomes part of the AI decision. The Bolt models are designed for RTX Spark PCs and DGX Spark workstations. If your team is due for a hardware refresh, which machine you get may now affect which AI tools you can use.
Agents may arrive for tasks you already do. Look at Nvidia's list of target tasks for Super VL: document intelligence, financial accounting and insurance claims. Those read like analyst and operations work. Because companies can post-train the model on their own data, your company could build an agent tuned to its own documents. Briski wrote that organizations can "own what they build." For you, that could mean an internal tool instead of a public chatbot.
Privacy is the selling point and the open question. TechCrunch points out that privacy-wary people may not want agents with deep access to information about their families, health and finances. Running locally keeps work off the cloud. The agent still needs access to whatever sits on your machine.
What to watch next:
- Bolt E4B's release later this month
- Which laptops from partners like Microsoft and Dell ship with RTX Spark
- Whether any outside testing of Nvidia's speed claims appears
- What your IT department says about local agents
Local does not mean pre-approved. Before you point a local agent at client files or financial data, check your company's AI and data policy. Rules written for cloud chatbots may not cover software that browses the web and calls tools from your own machine.
What we don't know yet
- Independent results. TechCrunch reports no independent tests of Super VL's speed or accuracy claims.
- Price and license terms. The report does not describe what the models or the RTX Spark PCs cost, or what terms govern their use.
- Real-world laptop performance. The report does not say how fast Bolt 74B-A9B runs on a typical AI PC for everyday work.
- Which apps will use Bolt. The report does not name the assistants or software that will run these models for ordinary users.
- How named customers use Super VL. Nvidia lists the companies, but the report does not describe what they use it for.
- Whether local agents ease privacy worries. TechCrunch raises the concern but does not say whether people will be persuaded.
Originally published at pickuma.com. Subscribe to the RSS or follow @pickuma.bsky.social for new reviews.
Top comments (0)