1X World Model
1X Technologies 路 September 2024
1X Technologies (1X AI Team)
World model 路 Closed weights 路 Also written 1XWM
- Parameters
- 14 B
- 14B for the January 2026 policy version, base model not named.
- GPU memory
- 26 GB
- weights at bf16, computed. Fits a 40 GB card
- Inference latency
- 11000 ms
- multiple GPUs, model not named, per 5 s of generated video; the inverse dynamics model adds about 1 s
- Weights
- closed
- closed.
What it is
The 1X World Model is the learned simulator 1X builds for its humanoids, first shown in September 2024 as an action-controllable video model trained on thousands of hours of EVE robot data. In June 2025 1X described it as a policy evaluator: it predicts future frames from candidate robot trajectories plus a value for the final frame, and with 70 percent accuracy it picks the better of two policies 90 percent of the time when their real success rates are 15 percent apart. Since January 2026 it also runs as a policy on NEO: a 14B video model generates 5 seconds of future video from a text prompt, and an inverse dynamics model turns the frames into actions. One such step takes about 11 s of generation plus 1 s for the actions, and the model is not public.
Architecture
- Backbone
- January 2026 version: a 14B text-conditioned video diffusion model, base not named. Earlier versions: not disclosed.
- Action head
- Inverse dynamics model (Depth Anything backbone with a flow matching head, two-image predictor, sliding window of 8 frames) that extracts actions from the generated video.
- Parameters
- 14B for the January 2026 policy version, base model not named. No size is published for the 2024 and June 2025 versions.
- Pretraining data
- 2024: thousands of hours of EVE humanoid data from homes and offices. January 2026 policy version: 900 hours of egocentric human video for mid-training, 70 hours of robot data for embodiment fine-tuning, and 400 hours of unfiltered robot data for the inverse dynamics model.
- Embodiments
- 1X EVE humanoid, 1X NEO humanoid
What hardware it needs
This is the section most people came for, so it is worth being precise about which numbers are measured and which are arithmetic.
| Precision | Weights | Note |
|---|---|---|
| fp32 | 52 GB | Training master weights, rarely used for inference |
| bf16 or fp16 | 26 GB | The usual way these checkpoints are served |
| int8 | 13 GB | Quantised, expect some loss of precision on fine motions |
| int4 | 6.5 GB | Aggressive quantisation, verify success rate before trusting it |
Smallest real card that fits the bf16 weights plus a 40 percent runtime allowance: 40 GB. That is an estimate for inference, not for fine tuning. Full fine tuning also holds optimiser state and gradients, which typically costs several times the weights unless you use LoRA or a comparable adapter.
| Inference latency, reported | 11000 ms on multiple GPUs, model not named, per 5 s of generated video; the inverse dynamics model adds about 1 s source |
Not published.
Reported results
Grouped by what the evaluation actually asked. Values from different suites are not comparable with each other, so each one keeps its suite and split.
Simulation
No results in this category are published for this model.
Real world
No results in this category are published for this model.
Fine tuned tasks
No results in this category are published for this model.
On real hardware
1X reports real-world runs on NEO with 30 trials per task for in-distribution and out-of-distribution tasks, but gives the success rates only in charts. Picking the best of eight generated videos improved success on a tissue-pulling task, according to the post.
Fine tuning it yourself
Not available. The public 2024 challenge offers a dataset and baseline world models, not this model.
Where it helps, where it does not
Strengths
- Used both to rank candidate policies and as a policy on the NEO humanoid.
- Its evaluator accuracy grows with data from autonomous policy rollouts, and predicted success rates correlate with real task scores, according to 1X.
- Learns largely from egocentric human video (900 hours) and uses 70 hours of robot data for embodiment fine-tuning.
Limits
- Closed: no weights or code for the model itself.
- Slow as a policy: about 11 s to generate 5 s of video plus 1 s to extract actions.
- Success rates are published only as charts, so they cannot be quoted as numbers.
- The NEO post-training data is mostly pick-and-place.
Sources
Everything on this page was taken from these documents. Where they disagree with what you read here, they win.
- https://1x.tech/discover/1x-world-model
- https://1x.tech/discover/redwood-ai-world-model
- https://1x.tech/discover/world-model-self-learning
Entry last checked 2026-11-12.
Try a policy on a real arm
A real SO-100 runs on /live and can be driven from the browser when it is online. Five of the models in this arena can be fine tuned on a dataset you record with your own arm.