CES week: Nvidia's open driving model, Atlas for factories, LTX-2

What shipped

Liquid AI

LFM2.5

A new generation of small on-device models from Liquid AI, trained for running on phones and laptops.

TII

Falcon-H1R 7B

Abu Dhabi's TII released a 7B reasoning model in its hybrid Falcon-H1 architecture family.

Lightricks

LTX-2

An open-weights video model that generates synchronized audio and video in one pass.

NVIDIA

Alpamayo

Nvidia opened a family of reasoning models, tools and data for autonomous driving, built around rare edge cases.

Boston Dynamics

New Atlas humanoid

Boston Dynamics showed a product version of Atlas built for industrial work, a step from lab demo toward factory floor.

NVIDIA

Vera Rubin platform

Nvidia said its next six-chip platform is in full production, with a claimed large cut in cost per token against Blackwell.

CES put robots and chips in the spotlight, but the model releases around it were worth the same attention.

Nvidia's physical AI push

The biggest story of the week was Nvidia's CES keynote. Two pieces matter. First, Alpamayo, a family of open reasoning models, simulation tools and datasets for autonomous driving. The pitch is that the hard part of self-driving is the long tail, the rare and strange situations, and that a model which can reason about those is worth opening up to the industry.

Second, Rubin. Nvidia said its next platform, a six-chip system built around a new Vera CPU and Rubin GPUs, is in full production, and it claims a big drop in inference cost compared with Blackwell. I take vendor cost claims with salt, but the direction is the same one I see every quarter.

I like the open part of Alpamayo for a practical reason. In regulated industries, the question is never "is the model smart." It is "can someone inspect it, test it, and sign for it." Weights and data that outsiders can examine are a better starting point than a black box.

Atlas heads toward work

Boston Dynamics showed a new Atlas aimed at industrial deployment. It is a humanoid, and the framing was about jobs in factories, not demos. I will believe the productivity story when I see numbers from a real shift, but the shift from research robot to product is the interesting part.

Small models and video

Liquid AI's LFM2.5 and TII's Falcon-H1R 7B both push capable models into sizes you can run locally. Lightricks opened LTX-2, a video model that produces audio and picture together, which cuts a step out of making short clips.

One more thing from the show floor. A lot of CES coverage was about gadgets with AI stuck on them, and I skipped most of it. The announcements I pay attention to are the ones with a clear buyer and a clear failure mode, like a driving model that can be audited or a robot built for a defined industrial task. Anything that depends on a slogan about intelligence, I file under later.

What I'd watch next: whether Boston Dynamics names a customer and a shift count, since that is what separates a launch from a deployment.