Zhiyuan GO-3 Preview Released: Learning Extends Past Deployment; Report Omits Weights and API
According to a Machine Heart (机器之心) report dated October 29, 2026, Zhiyuan released the embodied model GO-3 Preview, claiming first place on three benchmarks—RoboDojo, RoboTwin and RoboColiseum—and covering pre-training, post-training and post-deployment learning with WRAM, Genie Evolver and Genie Master. The report does not mention public weights, licensing or an external API.
What changed
According to a Machine Heart report dated October 29, 2026, Zhiyuan released the embodied model GO-3 Preview that day, claiming it ranks first on three benchmarks: RoboDojo, RoboTwin and RoboColiseum. The report divides learning into three stages: pre-training, post-training and post-deployment. On the pre-training side, it says about ninety percent of pre-training data comes from human first-person video. The model uses the WRAM architecture: CoAE extracts object positions, shapes and their relationships from camera frames; SVP predicts visual states at multiple future moments in a single forward pass, after which an inverse dynamics model combines them with the robot's own state to produce actions; KASO generates several candidate futures and selects those compatible with the real action for joint training. The report splits post-pre-training capability into zero-shot and few-shot tiers. Zero-shot means transfer without supervised fine-tuning on target-task real-robot data; few-shot uses a small number of demonstrations to improve precision and coordination on complex manipulation. It claims out-of-the-box capability on several hundred tasks. The cross-embodiment demonstration was on the bipedal humanoid X2, which opened a door while holding an object with both hands and picked up an object while kneeling. On the post-training side, Genie Evolver combines hundred-scale real robots, physics simulation and world models into a distributed environment, using CEL (collective experience learning) to aggregate execution experience. The figures given: at equal training time, as the number of participating real robots grows, success on a phone-packaging task rose from 50.8% to 98.3%, and failures in a 120-trial multi-stage long phone-packaging workflow fell from 25 to 2. On the post-deployment side, the agent framework Genie Master uses EEP to link capability invocation, execution feedback and replanning from breakpoints, and accumulates success, failure and exception-recovery cases as memory or training data. The report says the Jingling G2 running this framework has been deployment-validated in a real convenience store for restocking, picking and prepared-food packing.
What it means for developers
The report does not say whether GO-3 Preview's weights are public, what license applies, or whether an external API or SDK exists. From this source alone, developers cannot tell whether they can download or call it; the news is not a substitute for any existing integration path, and reads more as a signal about methods and evaluation. A few points are worth taking away. First, deployment-phase data flowing back into the policy means the model is not a checkpoint fixed at release: the same machine and the same instruction may land on different policies at different times, so assumptions about reproducibility and regression testing must change—this is an inference from the report's description, not something the source states. Second, the report clearly separates zero-shot from few-shot: zero-shot means transfer without supervised fine-tuning on target-task real-robot data, few-shot uses a few demonstrations to improve precision and coordination. That is an evaluation split developers can adopt directly, and is more actionable than a vague "out of the box." Third, cross-embodiment capability was demonstrated only on the X2, with two-handed door opening while carrying an object and kneeling pickup; the report gives no data for other arms and grippers, so embodiment-transfer cost remains unknown. Note that the three first-place benchmark results, the success rate and the failure counts are all vendor-reported. The source lists no per-benchmark scores, competing entrants, benchmark versions, task sets or random seeds, and the claim of reaching a win rate above ordinary human players "in a few hours" comes with no sample size or rule description. These are directional signals at best, not a basis for model selection.
Integration and evaluation
The following are recommendations and conditional analysis, not reported facts. First, whether to budget for demonstration collection. The report claims out-of-the-box capability on several hundred tasks and, after fine-tuning with a few demonstrations, completion of contact-heavy actions such as Lego assembly. If your tasks are mainly pick-and-place and your embodiment and demonstration setup are close to theirs, you can start with the zero-shot flow; if they involve insertion, deformable objects or high-precision alignment, prepare demonstration data in advance. The criterion: run the same task on the same embodiment at least fifty times each, compare the success-rate gap between zero-shot and few-shot, and require that gap to cover collection and labeling cost—otherwise fall back to zero-shot. Second, whether to invest in multiple robots for collective experience learning. The report says that at equal training time, as participating real robots scale up, phone-packaging success rose from 50.8% to 98.3%, via CEL retrieving successful segments, analyzing failure context and updating a shared policy. The precondition is multiple real robots that can collect in parallel, standardized tasks, and stable environment reset and data pipelines; with a single machine this benefit does not apply, and you should establish a single-machine baseline first. The source also does not say whether these robots must be homogeneous, which is an assumption to confirm with the vendor or verify yourself, not a known condition. The criterion: track failure counts in fixed trial batches (for example 120 per batch), require the cluster policy to keep improving over the single-machine baseline, and watch for regressions on existing tasks. Third, whether to allow online policy updates after deployment. Genie Master offers two paths in the report: experience entering prompts and memory to improve task planning, or execution and recovery data entering action-policy training. If tasks are irreversible or involve personal safety, enable only the memory layer first and keep a human fallback; if tasks are reversible and you have a shadow mode, then allow policy updates. The criterion: pass-rate change on the same frozen evaluation set before and after each update, and whether you can roll back to the previous version within an acceptable time.
Limits and open questions
GO-3 Preview is not currently in KimiSeek's model catalog, and the report mentions no external API availability. If an API or weight release appears later, availability should be judged by the vendor Zhiyuan's statements; this site's catalog is not evidence of obtainability. Whether weights are released and the license terms are two separate things, and the source addresses neither; "not mentioned" does not imply either "closed" or "open." The three first-place rankings are vendor-reported, with no opponent list, specific scores, benchmark versions, task sets or reproduction materials, so the margin and comparability cannot be judged, and the existence of public comparisons can be neither confirmed nor denied on this basis. The rise from 50.8% to 98.3% and the drop from 25 failures to 2 both come from vendor-provided experiments, with no disclosure of random seeds, failure criteria or degree of human intervention. "The industry's first validation of Physical RSI in a real business scenario" is a vendor statement without third-party verification, and the duration, number of stores and anomaly rate of the convenience-store deployment are not given. The report also provides no compute thresholds or cost information for training or deployment, and does not say whether the robots in the scale experiment were homogeneous.