← All model news

Baidu's Apollo World pitches a world model as a robotaxi examiner: 33 ms response, but no weights for developers

According to a Xinzhiyuan report dated November 4, 2026, Baidu Apollo released Apollo World, an agent-driven world model that it says can simulate trips of more than ten minutes continuously and respond within 33 milliseconds, and proposed a "2+3" standard for world models; the report does not mention weights, licensing, or an API.

Published Updated

What changed

According to a Xinzhiyuan report dated November 4, 2026, Baidu Apollo released a world model called Apollo World, which the report describes as "the first agent-driven, full-modality world model for physical AI." By its account, the system puts urban road networks, multimodal sensors, and traffic-agent interactions into a cloud simulation, supports real-time interaction at millisecond latency, and can drive continuously for more than ten minutes between arbitrary start and end points in a virtual city, completing a full trip from accepting an order through picking up and dropping off passengers. The report offers several technical details: every vehicle, rider, and pedestrian in the scene is modeled as an agent with independent perception and decision-making, so they react when the robotaxi changes its actions; the model jointly generates multi-view camera images and 3D lidar point clouds, and introduces a high-precision 3D road-structure prior to suppress long-horizon drift; and by distilling a multi-step diffusion model into a few-step student model, the report says real-time response is kept within 33 milliseconds (about 1/30 of a second). The report also says the model learned physical regularities such as lidar point-cloud attenuation in rain and longer braking distances on wet roads from more than 1 billion real frames, and that the next-generation L4 end-to-end AI Driver foundation model (ADFM) for Apollo Go has been deployed in vehicles, with over a hundred vehicles without safety drivers in operation. The report also relays Baidu's proposed "2+3" standard for world models: causality and interactivity as thresholds, and physical accuracy, persistence, and real-time performance as metrics for an L4 training ground. It should be emphasized that the performance figures, scene demonstrations, and deployment scale above are all one-sided statements from the report and the vendor, with no third-party reproduction or independent evaluation seen.

What it means for developers

The practical significance for developers has two layers. The first is methodology: the report distinguishes a "video generator" from an "interactive training ground" and lists causality, action-conditioned generation, long-horizon consistency, and real-time performance as checkable dimensions. Teams working on embodied AI, autonomous driving, or simulation data pipelines can use these dimensions as a checklist when building or procuring simulation environments, rather than directly trusting any one vendor's ranking. The second is integration: the report does not mention Apollo World's model weights, open-source license, API endpoints, billing, or a third-party application channel at all; it only says the system serves Apollo Go's own training loop. In other words, this is a public-facing technical release, not a model product that outside parties can call. There is currently no basis for treating it as a reproducible baseline or a directly integrable component. In addition, expressions in the report such as "only two in the world," "breaking the impossible triangle," and "AGI moment" are media and vendor rhetoric and cannot be taken as technical conclusions; figures such as 33 milliseconds, 100,000 kilometers per day, and obtaining a license in a few minutes also lack independent test-methodology descriptions. There have been several recent disclosure reports about Apollo World; what this piece adds is a description of nighttime testing in Shenzhen and a multi-destination continuous-trip demonstration in a virtual Hong Kong city, but it likewise does not resolve the question of external verifiability.

Integration and evaluation

The following are recommendations and conditional analysis, not reported facts. First, if you are evaluating closed-loop simulation training options, you can first break the report's "2+3" into measurable indicators: causality corresponds to whether the next frame can be inferred from historical frames alone without "knowing the ending in advance"; interactivity corresponds to whether the environment state changes after a given action; persistence corresponds to whether the map and lane structure drift after ten minutes of continuous simulation. The premise is that you have a runnable environment; if you cannot get Apollo World, you can only use these indicators to screen other simulators, and cannot claim on that basis that Baidu's approach is better. A suggested measurable protocol: run the same start and end points ten times and tally geometric drift and collision-judgment consistency. Second, if your workload is post-training of end-to-end driving models, the report says the value lies in amplifying rare hazardous events into large numbers of long-tail cases. You can first do a conditional check: use a short segment of real logs you collected yourself and see whether the simulation environment can reproduce the real vehicle's braking distance and point-cloud attenuation trends from the same initial state. If the pass rate is lower than expected, do not directly convert simulated mileage into real road-test mileage. Third, at the interface level there are currently no options. If what you need is a callable model, Apollo World is not in the selectable range; the ADFM mentioned in the report is likewise only described as an in-vehicle model, with no mention of external access. Do not infer from the report that a public API or weight download address exists.

Limits and open questions

First, neither Apollo World nor the ADFM version mentioned in the report currently appears in KimiSeek's live model catalog; if the report mentions any API or service availability, that is a vendor-side statement and unrelated to this platform. Second, the report does not state whether weights are released or under what license, nor whether there is a third-party integration or commercial licensing path; open weights and an open-source license are two different things, and neither can be determined here. Third, figures such as 33-millisecond response, 100,000 kilometers per day, more than 1 billion training frames, 28 cities, and 350 million kilometers of mileage are all one-sided claims, lacking descriptions of test environment, hardware configuration, and statistical methods. Fourth, the Shenzhen nighttime test and the virtual Hong Kong city demonstration are media impressions and cannot be treated as independent evaluation. Fifth, the physical AI market-size forecast is cited from Grand View Research and is an external institution's estimate; it does not constitute a judgment on Apollo World's commercial prospects.

Three minutes to your first working call

Create an account and an API key. Calls use your account balance. 30% off your first top-up. Top up ¥500, pay ¥350, receive ¥500 in credit. Top up ¥1000, pay ¥850, receive ¥1000 in credit You can still start with a ¥10 trial top-up.

Sources and editorial process

全球唯二,百度抢先交卷:首个物理AI世界模型发布

AI assisted the analysis and translation of the cited sources, followed by automated quality checks. Recommendations are not test results. Consult the current catalog for availability and pricing.