GEN-1
Generalist AI, United States · April 2026
- Parameters
- not published
- Not disclosed.
- GPU memory
- not published
- weights at bf16, computed
- Inference latency
- not published
- per action step
- Weights
- closed
- no public checkpoint
What it is
GEN-1 is Generalist AI's embodied foundation model, announced on 2 April 2026 as the successor to GEN-0 (November 2025). Its defining and most unusual design choice is that the pretraining set contains no robot data at all: it is built from over half a million hours of physical interaction recorded with low-cost wearable devices on humans, and roughly 99 percent of the parameters are trained from scratch rather than initialised from an existing VLM. Generalist reports that about one hour of robot data per task lifts average success from 64 percent (fine-tuned GEN-0) to 99 percent, with roughly 3x faster execution. There is no paper, no parameter count, no architecture description and no standard benchmark, so every number rests on the company's own blog.
Architecture
- Backbone
- Not disclosed. Generalist states approximately 99 percent of the parameters are trained from scratch rather than fine-tuned from an existing model, and that architecture, training and inference were all designed in-house.
- Action head
- Not disclosed. Described only as a large multimodal model that emits actions in real time, with an inference-time technique the company calls Harmonic Reasoning and does not explain.
- Parameters
- Not disclosed. The only architecture claim is that about 99 percent of parameters are trained from scratch.
- Pretraining data
- Over half a million hours of high-fidelity physical interaction data. Generalist states the pretraining dataset contains no robot data and instead uses recordings from low-cost wearable devices on humans performing millions of activities.
- Embodiments
- not named in the source; Generalist states GEN-1 adapts to a new robot embodiment and a new task simultaneously, a July 2026 follow-up post describes support for a range of end effectors
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.
Not published. No hardware, accelerator or memory requirement appears in any primary source.
The authors publish no memory or latency figure for this model. Everything above is computed from the parameter count.
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
Rollouts on physical hardware. The setups differ, so read these as evidence, not as a ranking.
- Generalist AI internal real-robot evaluation aggregate over the reported tasks, about 1 hour of robot data per task99%success rateCompany-reported, no external benchmark. Comparison points in the same post: fine-tuned GEN-0 64 percent, trained from scratch without pretraining 19 percent.source
- Generalist AI internal real-robot evaluation robot vacuum servicing99%
- Generalist AI internal real-robot evaluation box folding99%
- Generalist AI internal real-robot evaluation packing a phone into a case, cycle time15.5 stask completion timeCompany-reported, 2.8 times the speed of GEN-0. Task duration, not model latency.source
- Generalist AI internal real-robot evaluation box folding, cycle time12 stask completion timeCompany-reported as about 12 seconds, a 2.8x speedup. This is task duration, not model latency.source
Fine tuned tasks
No results in this category are published for this model.
On real hardware
All reported results are real-robot task evaluations run by Generalist. Robot vacuum servicing rises from 50 percent (GEN-0) to 99 percent. Box folding rises from 81 percent (GEN-0) to 99 percent and takes about 12 seconds, 2.8 times faster. Packing a phone into a case takes 15.5 seconds, 2.8 times the speed of GEN-0. Training from scratch without pretraining averages 19 percent on the same tasks.
Fine tuning it yourself
Generalist states each published result was produced with approximately one hour of robot data for that task, and that adapting to a new task means adapting to that robot embodiment for the first time as well, because the base model saw no robot data. No public tooling, data format or access path exists beyond a partnership inquiry.
Where it helps, where it does not
Strengths
- Pretrained without any robot data, which decouples the corpus from robot fleet ownership and is the only approach of its kind in this comparison.
- Over half a million hours of physical interaction data, the largest disclosed pretraining corpus among all models here.
- About one hour of robot data per task is enough to reach the reported success rates, the lowest published adaptation cost in this set.
- Reports explicit generation over generation and from-scratch baselines (99 versus 64 versus 19 percent) rather than absolute numbers alone.
- Speed is treated as a first-class metric, with published cycle times rather than success rate alone.
Limits
- No paper, no parameter count, no architecture description, no latency figure and no hardware requirement is published anywhere.
- Every number comes from Generalist AI's own blog. There is no standard benchmark such as LIBERO or SimplerEnv and no third-party evaluation.
- The robot embodiments behind the reported results are not named, so the results cannot be reproduced or scoped to specific hardware.
- Generalist states that not all attempted tasks reach these rates, and that some tasks would require even higher success rates or speeds to be useful in real settings.
- Harmonic Reasoning is named as an inference-time technique but never defined.
- Weights and code are closed, access is by partnership inquiry only.
Sources
Everything on this page was taken from these documents. Where they disagree with what you read here, they win.
- https://generalistai.com/blog/gen-1
- https://generalistai.com/blog/beyond-world-models
- https://generalistai.com/blog
Entry last checked 2026-08-11.
Try a policy on a real arm
A physical SO-100 is online and free to drive in the browser, and five of the models in this arena can be fine tuned on a dataset you record with your own.