1X puts a 14-billion-parameter video model in charge of its NEO home humanoid, trading robot data for internet video

1X Technologies said on January 12, 2026 that NEO, the home humanoid it began taking orders for in October 2025, would be driven by a new policy built on a generative video model rather than a conventional robot control network. The company calls it the 1X World Model, or 1XWM. In its announcement chief executive Bernt Børnich said NEO could now learn from internet-scale video and carry that knowledge into the physical world, and 1X researcher Daniel Ho said a user's prompt could become a fully autonomous action even for tasks the robot had never seen.

The technical write-up 1X published the same day gives the numbers behind that claim. The backbone is a 14-billion-parameter, text-conditioned diffusion model for video. It was trained in three stages, starting with web-scale video, then mid-training on 900 hours of egocentric human video of people manipulating objects, then fine-tuning on 70 hours of NEO sensorimotor logs. A separate inverse dynamics model, trained on 400 hours of unfiltered robot data that includes random play, reads the generated frames and predicts the actions needed to move from one frame to the next. It uses a Depth Anything backbone with a flow matching head and works over a sliding window of eight frames.

In practice NEO receives a text prompt, the world model renders five seconds of predicted video, and the inverse dynamics model extracts the motor commands. 1X said generation takes 11 seconds on multi-GPU inference and action extraction one second, so each five-second action chunk costs roughly 12 seconds of compute before the robot moves. The company also uses a vision-language model to expand short prompts into detailed captions, a step it calls caption upsampling, which it says improves how closely the video follows instructions.

The company was unusually open about results. In real-world tests of 30 trials per task, 1X charted success of about 100% for grabbing a bag of chips, about 80% for steaming a shirt, about 60% for opening a sliding door and for using a watering can, and about 40% for ironing a shirt and scrubbing a dish. Pouring cereal and drawing a smiley face both scored zero. On a tissue-pulling task, a single generation succeeded 30% of the time, and picking the best of eight parallel generations lifted that to 45%. 1X listed its own limits as weak three-dimensional grounding from monocular pretraining, a tendency to overshoot or undershoot positions even when the video looks right, difficulty with dexterous tasks and a five-second task horizon.

The launch sits on top of a consumer product with firm commercial terms. 1X, headquartered in Palo Alto, California, opened NEO orders on October 27, 2025 at $20,000 for early access with priority 2026 delivery, or $499 a month on subscription, with US deliveries in 2026 and other markets from 2027. Its launch release put NEO at 66 pounds with a lifting capacity of 154 pounds, a carrying capacity of 55 pounds and a noise level of 22 decibels, and offered an Expert Mode in which 1X staff can guide the robot through tasks it cannot yet do alone.

The design choice puts 1X on a different path from most of its rivals. Its write-up contrasts 1XWM with vision-language-action models such as Physical Intelligence's π0.6, Figure's Helix and NVIDIA's GR00T N1.5, arguing that models built on vision-language backbones learn semantics well but are not trained to predict physical dynamics. A video model, in 1X's framing, already knows how objects move and fall, so the robot needs far less of its own data. That is the point of the 70-hour fine-tune, which is small next to the thousands of teleoperated hours most humanoid developers describe collecting.

The money behind the competing approaches is large. MEKAWIRE company data show Skild AI closing a $1.4 billion Series C at a $14 billion valuation two days later, on January 14, 2026, and Figure raising $1 billion at a $39 billion post-money valuation in September 2025. The latest 1X round in the same records is a $100 million Series B from January 2024. NVIDIA moved in the same direction two months after 1X, previewing GR00T N2 at GTC in March 2026 as a model that predicts video before acting.

The open questions are speed and reliability. A 12-second wait for each five seconds of motion is slow for a robot sharing a kitchen with people, and 40% success on ironing or dishwashing is far from what a $20,000 appliance will be judged against. 1X said its next steps are faster inference for reactive and contact-rich work, closed-loop replanning with memory, depth or stereo input for better 3D grounding, and a self-improvement loop driven by the robot's own experience rather than expert demonstrations. Expert Mode remains the fallback, which means remote human operators stay in the loop for an unknown share of tasks.

Later disclosures show how the bet developed. On April 30, 2026 1X said it had opened a 58,000-square-foot factory in Hayward, California with capacity for 10,000 NEOs a year and a goal of more than 100,000 a year by the end of 2027, and on June 4, 2026 it set up a 1X World Model Lab under Sam Sinha, formerly a founding research scientist at Luma AI, to scale world model pretraining.