Nvidia launches new open models for AI agents

Nvidia wants AI agents running on your own hardware

Nvidia wants AI assistants to run directly on your computer, with less dependence on the cloud.

The company has introduced three new open-weight models, including the 120-billion-parameter Nemotron 3.5 Super VL for enterprise agents and two smaller Bolt models designed for local PCs and workstations.

These models can handle tasks like document analysis, financial accounting, and web browsing, while allowing companies to customize them using their own data.

The local PC models are especially interesting.

I think Nvidia sees an opportunity in something AI companies haven't fully solved: convincing people to give assistants access to their personal information. Running an agent locally could make that decision easier, particularly when it involves private documents, financial records, or sensitive work files sitting on someone's laptop.

There's a business incentive, too. If more AI runs locally, Nvidia has another reason to sell its chips through PC makers like Dell and Microsoft.

The real question is whether these smaller models will perform well enough for people to stop relying on cloud-based assistants.

Wabi thinks AI assistants should work with groups

Twelve years after pitching an AI concierge, Replika founder Eugenia Kuyda is giving the idea another shot.

Her new app, Wabi, launched publicly with a different approach to personal AI assistants. It organizes tasks into separate conversations, automatically builds custom apps, and lets people share AI-powered group chats with partners or family.

The interesting part is the group chat.

Imagine asking an assistant to check your partner’s calendar, track concert tickets, or find nearby events for your kids without switching between five different apps.

I think Wabi is onto something because most AI assistants still assume we're working alone, even though plenty of everyday decisions involve other people. Letting AI handle shared plans could make it considerably more useful, especially for families managing busy schedules.

But there's a catch. Wabi can access sensitive information, including email, calendars, and potentially bank details, making privacy controls especially important.

Would you trust an AI assistant with information your partner could ask about?

AI can pass medical exams but struggles to finish the work

AI models can score nearly perfectly on medical exams, yet Stanford researchers found that even the best agents struggle to complete actual healthcare tasks.

In a new study of clinical workflows, the top-performing model completed just 46% of tasks, while the best agent handling administrative work managed 36.3%.

One finding stood out. In many cases, AI correctly documented what needed to happen but failed to actually place the medical order in the electronic health record.

That's concerning.

I think this exposes a serious problem with how we judge AI in healthcare, where a well-written response can look convincing even when the required action never happens. Hospitals need systems that reliably complete workflows involving patient records, insurance portals, and documentation, especially when a missed step can delay treatment.

The research also suggests that better memory and coordination between tasks could matter as much as more powerful models.

If AI can explain the right treatment but forget to order it, how much responsibility should hospitals give it?

Prohuman team

Free Guides

Explore our free guides and products to get into AI and master it.

All of them are free to access and would stay free for you.

Feeling generous?

You know someone who loves breakthroughs as much as you do.

Share The Prohuman it’s how smart people stay one update ahead.