Nvidia just dropped three open Nemotron models, and the biggest one is a decoy. The 120-billion-parameter Nemotron 3.5 Super VL is impressive — it handles documents, images, and multistep tool use — but it's not the story. The story is the two Bolt models: a 74B-A9B variant squeezed into a 24GB footprint and a 4GB E4B model for phones and robots. That's like Ferrari releasing a family sedan that gets 100 miles per gallon. It changes who can play the game.
Here's why 24GB matters. That's exactly the VRAM on a single consumer RTX card. Nvidia isn't just shipping a model; it's shipping a reason to buy the hardware it already sold you. The pitch is simple: run a capable agent locally, no cloud bill, no data leaving your machine. For regulated industries like healthcare and finance, that's not a feature — it's the whole ballgame.
But here's the twist: Nvidia's biggest customers are Microsoft, Amazon, and Google. They built their AI empires on metered token inference in their data centers. A free, open model that runs multistep agent tasks on your desk undercuts that. So is Nvidia betraying its own cloud partners? Not exactly. Nvidia wins either way. If inference stays in the cloud, it sells rack-scale GPUs. If inference moves to the desk, it sells RTX cards and DGX Spark boxes. No competitor has that dual position. AMD and Intel would love to own the local-agent story, but they lack Nvidia's software gravity. Cloud providers can't credibly argue against privacy and latency.
The clearest losers are inference API providers — the Together/Fireworks tier — and any startup whose entire pitch is cheap hosted inference for agent workloads. A free, open, 24GB model on hardware the customer already owns is a brutal competitor for metered API calls. That's like Netflix suddenly giving away Blu-rays.
Also worth noting: Nvidia quietly dropped the 'Llama' from its branding. Earlier models were 'Llama Nemotron,' built on Meta's weights. The 3.x family is plain 'Nemotron,' with Nvidia's own architecture. Nvidia graduated from repackaging Meta's work to fielding an in-house family. Meta still partners with Nvidia on training hardware, but now they compete for the default open enterprise model. That's a relationship to watch.
The Bolt E4B model, scheduled for October 23 on GeForce RTX, Jetson, and RTX/DGX Spark hardware, could make always-on local agents as common as a webcam. But privacy still depends on system configuration and how connected tools handle data. And businesses still need to test accuracy and resource use. So don't throw away your cloud API keys yet. But if you're building an agent that handles sensitive data, you might want to clear some space on your desktop. The future of AI might not be in the cloud. It might be under your desk.