ACE Robotics Open-Sources Kairos 3.0 On-Device World Model — 4B Parameters Leave Cosmos 2.5 Behind by 72×

On June 15, ACE Robotics officially announced the completion of its Angel+ round of financing, only four months after the previous round. The total financing this year has already reached hundreds of millions of USD. The next day, the company released its Kairos 3.0 world model — which had just achieved SOTA results on four global embodied intelligence leaderboards — and, even more impressively, simultaneously open-sourced the on-device version with 4B parameters, available at github.com/kairos-agi/kairos-sensenova.

For a company founded only in July 2025, this pace is far from slow — it even feels a bit like “the king of hustle.”

In one sentence

Kairos 3.0-4B achieves three feats no one else has done:

• World’s first on-device world model that can directly drive the physical body of a robot: Runs on Jetson Thor T5000 (517 TFLOPS) and can output complete control commands from upper limbs to fingers to lower limbs, bypassing intermediate translation;
• Industry’s first to achieve 1:1.5 real-time generation on the THOR platform (video generation time vs. video length), while cloud-side achieves 1:1 real-time;
• Generates a 10-second clip in just 9.5 seconds on A800, ~72× faster than NVIDIA Cosmos 2.5 (687.2 seconds), 9× faster than Alibaba Wan 2.2, and 151× faster than Ant Lingbot.

First place in all four benchmarks — RoboTwin 2.0, LIBERO-Plus, WorldModelBench Robot, DreamGen. On PAI-Bench-robot it scored 80.03, surpassing Cosmos 2.5-14B (79.4) and Lingbot (79.96); in DreamGen Bench, the physical alignment metric PA hit 0.529, nearly 70% higher than Wan 2.2-5B’s 0.314.

This is a watershed moment: in the world model track, the bonus period of parameter stacking is over — architecture-native design + inference engineering optimization is reshuffling the field.

Why “4B beats 14B” is not hype

Over the past year, the mainstream approach in the industry to world models has been “big model retrofit” — take a large video-generation model (Cosmos, Wan series), attach a motion head, and roughly let the robot move. The core problem: video models lack causal priors of the physical world — speed of flowing water, rigidity of stones, support relationships — everything is guessed from data.

Kairos 3.0 follows a different path: designing from the architecture’s foundation for real-world robot operation, using the laws of nature and causal chains as cognitive basics. Officially described as a “multi-modal understanding—generation—prediction” integrated architecture. Sounds lofty, but differences are striking in scenarios:

• Pouring water task: Kairos generates steady flow speed; liquid amount strictly matches cup volume, in line with mass conservation; Cosmos 2.5 and Lingbot either flow too fast or magically increase liquid volume, breaking physical logic;
• Stacking balancing stones: Kairos obeys gravity and structural support; Cosmos’ stones begin floating, Lingbot’s bottom stone even disappears out of nowhere.

This is not benchmark number play — it’s about whether a world model truly “understands” physics. A “performer” world model that doesn’t actually work is an accident generator in real scenarios.

Inference acceleration comes from an engineering trick: self-developed hybrid temporal linear attention operator. Regular Transformers’ quadratic complexity is disastrous for long video generation; hybrid temporal linear attention cuts this overhead, and with 4B lightweight parameters (23.5GB VRAM, same as Wan 2.2-5B, far below Cosmos 14B’s 70.2GB), it can fit into devices like Jetson Thor.

What on-device real-time generation means

In the embodied industry, there has been a long-standing embarrassment: cloud inference is fast but suffers from latency/bandwidth limits for real-time control; on-device models that can run are too “dumb.” The result: robots look smart in demo videos, but in reality either stutter or make mistakes.

Kairos 3.0-4B’s 1:1.5 real-time generation on Jetson Thor T5000 is a key milestone. In practice, for a 10-second action, the model needs only 15 seconds to predict, plan, and send the complete control sequence. Combined with cloud’s 1:1 real-time inference, the full “prediction—decision—execution” loop can be closed independently from the cloud for the first time.

More importantly, this on-device version is fully open-source — a first in China’s embodied field. Previously, offerings were either closed SDKs or stripped demo weights. With Kairos 3.0-4B releasing full on-device capabilities, small teams, university labs, and second-tier robot makers can run it directly without being bottlenecked by compute power or licensing.

One long video — “long temporal sequence” up to 7 minutes

Another hard problem for world models is long temporal degradation — beyond a few dozen seconds, coherence breaks. Kairos 3.0 uses agent technology for hierarchical task decomposition and self-reflection mechanisms, producing coherent embodied interaction videos up to 7 minutes.

In the official home-scene demo, the robot completes tidying the table, laundry, and breakfast preparation in one continuous take, with no breaks. This duration has real industrial significance: most assembly line processes, hotel cleaning routes, and warehouse sorting workflows are in the minutes range — Kairos’ long temporal capability makes “replacing part of real-machine training with world model simulation” technically feasible.

Data paradigm: why Wang Xiaogang insists on “human-centered”

Talking Kairos means talking ACE’s R&D paradigm. Chairman Wang Xiaogang (co-founder of SenseTime) repeatedly stresses: data shortage in embodied intelligence is cliff-like.

Some striking numbers:

• In autonomous driving, Tesla FSD V14 uses its world model simulator to train at the daily equivalent of 4 million hours of human driving — about 500 years;
• The total industry embodied intelligence real-machine data volume is only about 100,000 hours.

The scale gap means sticking to real-machine teleoperation can’t catch up. Switching entirely to pure vision learning (Figure, Tesla recently trying) hits the “reality gap” — video can’t see force, friction, or 3D mechanics.

ACE’s solution: “human-centered” environment capture — using multi-modal devices across perspectives to record human–environment interactions, combining first/third-person video, force/tactile data, motion trajectories, and audio to build a physics-grounded 3D asset library. This expanded training data to 1 million hours, 10× traditional real-machine capture. With Kairos’ generation amplified, effective data scale could reach billions of hours.

This differs from Figure and Tesla’s “hardware—data—model” closed-loop path. Wang admits China’s closed loop isn’t formed yet, but he bets on: world model as amplifier, enabling data efficiency to catch up nonlinearly.

Cross-body generalization — train once, multiple embodiments

Traditional embodied models suffer from “one body, one training” — change the robot, redo all data and strategy. Kairos 3.0-4B supports cross-body one-click generation — single arm, dual arm, dexterous hand — without extra training, outputting execution strategies.

Already adapted hardware includes mainstream models like Zhiyuan-Elf G1 and Unitree G1. In ACE’s own deployment pace, six months ago the embodied brain module A1 mainly served road inspection robot dogs; now it’s in hotels, unmanned retail, unmanned logistics warehouses. Forms expanded from robot dogs to wheeled, bipedal, and robotic arms — powered by Kairos’ cross-body generalization.

Domestic chip ecosystem — a detail to note

Kairos 3.0 has been adapted to multiple domestic chips from companies like Moore Threads, Biren Technology, Sugon, Huixi Intelligence, Inno-Sensor. Moore Threads is also an investor in this Angel+ round.

This setup recalls the DeepSeek + domestic chips playbook: model side does engineering optimization to squeeze performance from domestic chips; chip side gets top-tier models as benchmark workloads. For China’s embodied track, constrained by NVIDIA compute, this is a necessary path.

Horizontal comparison — who it’s against and differences

Putting current competitors in one table:

| Model | Parameters | VRAM | 10s generation time | On-device deployment | |-------|-----------|------|---------------------|----------------------| | Kairos 3.0-4B | 4B | 23.5GB | 9.5s | ✅ THOR 1:1.5 | | Cosmos 2.5-2B | 2B | — | 687s | ❌ | | Cosmos 2.5-14B | 14B | 70.2GB | — | ❌ | | Wan 2.2-5B | 5B | ~23GB | 85s | ❌ | | Lingbot | 28B | 46.1GB | 1436s | ❌ |

Kairos’ advantages are clear: small parameters, low VRAM, fast inference, on-device capability. But some concerns warrant further observation:

• Breadth of physics simulation coverage: Demos show pouring water, stacking stones, home scenes — industrial repetitive tasks, precision assembly, flexible materials handling need more data validation;
• True generalization boundaries for cross-body: Mainstream models like Zhiyuan, Unitree are fine — long-tail adaptation costs aren’t public;
• Capabilities gap between open-source and commercial versions: 4B is open-sourced, but unclear if larger models are available for commercial deployment.

Final note

In 2026’s embodied track, financing numbers are massive, but few deliver real substance. The most notable part of ACE’s move isn’t “four SOTAs” or “72× speedup” marketing — but two conclusions:

First, the paradigm contest for world models in embodied intelligence is decided — architecture-native designs are outperforming “big model retrofits.” Players solely stacking parameters on video generation models will struggle in the next 1–2 years.

Second, on-device real-time is the tipping point for embodiment. Before Kairos 3.0-4B, “can robots think independently, without the cloud” was an open question; after — it’s an engineering question.

How far open-source can go depends on the GitHub repo’s issues and PRs. But at least ACE chose a more difficult yet higher-ecosystem-potential path than closed SaaS.

References

• Kairos-Sensenova Open-Source Repository (GitHub): Kairos 3.0-4B on-device version code and weights release address