Zhiyuan releases GO-3 Preview and WITA-OMNI dual models, and proposes the Worken work-hour unit

Zhiyuan actually came up with a way to give performance reviews to the robots doing the work (bushi).

At this year's Innovation Day, Zhiyuan co-founder, president and CTO Peng Zhihui (Zhihuijun) unveiled a new physical AI unit called "Worken," used to measure a robot's effective intelligent work hours.

Zhiyuan also designed a performance settlement formula for Worken:

Worken = number of active robots × autonomous working hours × output efficiency ratio

The number of active robots reflects deployment scale, autonomous working hours look at whether a robot can keep working without human takeover, and the output efficiency ratio measures actual output efficiency.

Fine, from now on silicon-based beings can no longer just slack off ╮(╯▽╰)╭

But this time Zhiyuan isn't just arranging performance reviews for robots.

Around how to make robots truly create value, Zhiyuan also brought a series of technical advances spanning models, bodies and operating systems, covering core embodied intelligence capabilities such as motion, operation and interaction.

For instance, on the new model side, the embodied foundation model GO-3 Preview has already taken first place on three embodied intelligence evaluation leaderboards, and the interactive large model WITA-OMNI also topped four evaluation leaderboards.

But for robots to keep delivering this performance, a set of calculation formulas alone isn't enough.

The new models, new bodies and new systems Zhiyuan released this time happen to string together the technical layout behind Worken.

From models to bodies, what exactly did Zhiyuan release?

At this Innovation Day, Zhiyuan released the new-generation embodied foundation model GO-3 Preview and the omni-modal end-to-end interactive large model WITA-OMNI.

The two models target two core categories of robot working capability respectively.

GO-3 mainly targets operational intelligence, helping robots understand tasks, generate actions and complete operations; WITA-OMNI targets interactive intelligence, enabling robots to understand people, respond to people, and communicate naturally in real environments through voice, facial expressions and actions.

At the same time, Zhiyuan is also advancing upgrades to motion intelligence, hardware architecture and the operating system.

Models are responsible for improving capabilities, bodies and systems are responsible for carrying and coordinating capabilities, and real deployment tests whether these technologies can play a role on the work site.

Thus came Zhiyuan's "three intelligences" layout of interactive intelligence, motion intelligence and operational intelligence.

Operational intelligence

First, operational intelligence: Zhiyuan's latest contender is the embodied foundation model GO-3 Preview.

It took the crown on three embodied intelligence evaluation leaderboards—RoboDojo, RoboTwin and RoboColiseum—completing a sweep of all three.

It is reported that GO-3 Preview adopts the embodied native architecture WRAM (World Reasoning Action Model), combining task understanding, reasoning, future state prediction and action generation.

When robots perform fine operations, they must understand task goals and plan actions, and also predict the changes the actions may bring. WRAM attempts to bring these links into a unified learning framework to improve the model's ability to handle complex tasks.

In the pre-training stage, based on the joint scaling of models and data, GO-3 Preview has stronger task transfer and complex skill adaptation capabilities, thus performing better in Zero-Shot and Few-Shot tasks.

No robot can escape the Lego test!

After fine-tuning, a robot equipped with GO-3 Preview can complete fine operations such as part grasping, alignment and insertion, and complete assembly tasks under Agent orchestration.

In addition, the model also supports cross-body transfer. Switched onto the bipedal humanoid robot X2, it can also open doors with both hands and kneel down to pick up objects.

But as you know, the laboratory environment is relatively simple, while in reality the environment and object positions change endlessly, and embodied models cannot learn all their skills through pre-training alone.

To this end, Zhiyuan built the post-training system Genie Evolver, combining a hundred real machines, physical simulation and world models—equivalent to equipping robots with an internship workshop.

Once they enter actual positions, the embodied agent framework Genie Master continues to take on task planning and exception recovery, turning today's mistakes into tomorrow's experience.

General capabilities rely on training—improvement through virtual-real interaction—real experience reused again; linked together in one chain, embodied robots are completely different from traditional equipment.

Interactive intelligence

If GO-3 focuses on how robots complete work tasks, WITA-OMNI focuses on how robots communicate with people.

As Zhiyuan's newly released interactive large model, WITA-OMNI took first place in all four authoritative evaluations: Daily-Omni, TimeLens-Bench, EmbSpatialBench and MMAU.

WITA-OMNI jointly models input speech, vision and text together with output speech, facial expressions and actions in the same model, enabling robots to listen, see, speak and move at the same time.

For example, in open human social environments such as hotels and scenic spots, robots need to determine in noisy surroundings who is talking to them, whether the other party has finished speaking, and when it is appropriate to respond.

One quite interesting feature is that WITA-OMNI does not need a wake word.

Zhiyuan enables the model to integrate information such as sound direction, voiceprints, faces and lip movements to identify the conversation partner, and then combine gaze and tone to judge the timing of interaction. Even if the robot is performing other actions, it can continue to participate in the conversation.

OK, back to the main topic!

WITA-OMNI adopts a fast-and-slow dual-system design.

The front-end Interaction Engine is responsible for immediate responses.

The back-end Background Agent is responsible for reasoning, planning, tool invocation and long tasks.

When encountering complex problems, the back end continues processing while the front end can still maintain the conversation and synchronize progress, and users can also add requirements or adjust tasks at any time.

In the on-site demonstration, Xiong Yan, Zhiyuan partner, co-president and president of the interaction business system, fired off a rapid series of questions, and the robot equipped with WITA-OMNI answered fluently; when asked "who is the most handsome at Zhiyuan," it took the opportunity to be a bit narcissistic.

According to data disclosed by Zhiyuan, WITA-OMNI can usually respond within an average of 600 milliseconds, and as fast as about 400 milliseconds. This speed is already close to the natural reaction in human conversation.

In terms of language, WITA-OMNI supports understanding more than 100 languages and multiple dialects, can speak 29 languages and 7 dialects, and supports local interaction without a network.

Another interesting feature is Intelligent FaceUI.

Users only need to ask a question, for example, "explain the Pythagorean theorem to me," and the robot can dynamically present information on its facial screen.

When encountering complex problems that are difficult to explain clearly for the moment, graphical expression is obviously much more intuitive than dry voice announcements.

Zhiyuan plans to complete full deployment in China before December 2026, launch an overseas version in Europe in January 2027, and complete full launch in overseas regions in March.

Motion Intelligence

In addition to work intelligence and interaction intelligence, Zhiyuan is also continuously advancing innovation in motion intelligence.

The AGILE general motion intelligence framework released on Innovation Day incorporates environmental perception, terrain understanding, whole-body motion control and coordinated upper-limb and lower-limb operation into a unified perception and control loop.

Combined with an autonomous evolution architecture, motion intelligence, as foundational intelligence, moves from motion control toward spatially mobile intelligence capable of autonomous evolution, better meeting the spatial mobility requirements of the deployment state.

When facing dynamic obstacles, multi-person collaboration and unexpected human-robot interaction, robots need to adjust their actions according to environmental changes while also taking safety into account.

New Architecture and New Body

In the early days of the industry, most robots only had a single intelligent capability, and later many solutions moved toward the separation of the three intelligences, that is, "three sets of intelligence running on separate hardware boards," which is also the route many manufacturers are currently trying.

This time Zhiyuan goes a step further by launching the third-generation hardware architecture AHA3, bringing the three sets of intelligence—motion, interaction and work—into the same control system.

AHA3 is uniformly scheduled by a single SoC, achieving the integration of the three intelligences.

According to data disclosed by Zhiyuan, this upgrade reduces physical volume by 50% and improves response speed by 90%.

And according to the calculation method of the performance assessment formula, all of a robot's intelligence must ultimately be cashed out as Worken through a reliable body.

Dei, Zhiyuan also released a new robot body: the safety-standard heavy-load embodied robot Genie G2 Max.

It supports a single-arm payload of 20 kilograms and a dual-arm peak payload of 50 kilograms, is equipped with seven-degree-of-freedom force-controlled dual arms, and through designs such as multiple safety protections and dual-battery hot swapping, it targets long-duration continuous operation needs.

Also released on the same stage were the world's first subcutaneous joint liquid cooling system and a new generation of dexterous hand family, further solidifying the foundation of the body and core components.

Using the embodied native OS to uniformly schedule robot software and hardware

As models and robot bodies continue to upgrade, how to make different hardware, underlying control and upper-layer intelligent applications work together efficiently has become another engineering problem.

This embodied intelligence native operating system covers the entire process from construction and deployment to operation, handling hardware access and operation at the lower level and providing an application runtime environment for motion intelligence, work intelligence and interaction intelligence at the upper level, and it has already been verified by mass-produced robots.

One cannot say the pressure is not enormous...

Zhiyuan says that Lingqu OS first solves the problem of hardware access.

Through a unified hardware abstraction framework, it supports rapid access for 15 categories and more than 100 types of peripherals, and is compatible with mainstream computing power platforms.

For hardware developers, new devices can be connected to the system faster, reducing the workload of repeated adaptation and also leaving room for different hardware manufacturers to innovate.

Another key point is real-time performance.

When robots perform high-frequency motion control tasks, the system must not only respond quickly but also ensure that response time is sufficiently stable. That is, the system needs to minimize the time fluctuation each time an instruction is executed, so that the robot completes actions as expected.

Lingqu OS has optimized deterministic latency across the full stack, and under a 1000Hz high-frequency task load, the end-to-end latency jitter is less than 4 microseconds.

On this basis, Lingqu OS also provides an intelligent service framework, supports mainstream industry inference frameworks, and encapsulates system services as reusable Skills.

Developers can directly call the ready-made capabilities (Skills) encapsulated by the system to build applications, without having to start adapting from the bottom layer for everything.

Moreover, Zhiyuan has put this thing into open source service!

By the end of this year, Zhiyuan will fully open source Lingqu OS 1.0 and donate the project to the OpenAtom Open Source Foundation.

According to the plan disclosed at the main forum, Zhiyuan will also open source 2 billion lines of code that have been validated through mass production, more than 5,000 ecosystem software packages, provide an end-to-end construction toolchain, and pre-integrate a variety of embodied intelligence open source components.

Friends who have not used products from Zhiyuan's own system can wait until then to experience whether Lingqu is easy to use with their own software and hardware.

No matter how completely the technical roadmap is explained, in the end it still comes down to whether the robots can keep working continuously at the job site.

Now, Zhiyuan has already deployed robots in scenarios such as industry, retail, and culture and tourism, and different scenarios have different requirements for them.

In industrial scenarios, what matters most is task precision, work cycle time, and operational stability.

At present, Zhiyuan has partnered with Longcheer (a leading global consumer electronics ODM manufacturer) to build the world's first complete 3C production line covered by embodied intelligence.

This production line achieved large-scale production in March 2026, and as of the convening of Innovation Day, it had been running stably for more than 7 months.

In a previous 6-day full production line livestream, the robots completed more than 64,824 operations in total, produced 17,625 tablet computers, and achieved a operation success rate of 99.99%.

Their average work cycle time per item is 22 seconds, on par with skilled human workers.

Next, look at the retail scenario, which is the most demanding test of interactive understanding and task changes.

In Aishida's offline stores, 100 Lingxi X2 robots served as silicon-based shopping guides, driving a total of 2.23 million yuan in sales over 3 days.

The culture and tourism scenario, by contrast, has to face constantly changing foot traffic, noisy environments, and multilingual communication, and must also continue operating during daily business hours.

Hengqin Chimelong Spaceship Park has deployed nearly 500 robots, covering more than 100 positions and undertaking services such as explanation and interaction, receiving about 10,000 people per day on average during the National Day period.

Now, when you go to the Whale Wonder World exhibition area, robots can already use WITA-OMNI to generate real-time explanation visuals, putting you directly at the forefront of the intelligent experience (doge).

The October Worken dashboard shown at the main forum listed data from scenarios such as the Longcheer production line, Guangzhou Metro, Hengqin Chimelong, and Aishida retail stores.

What a sigh of emotion: embodied intelligence has been charging forward like a runaway horse for so long, and whether in deployment scale, autonomous working hours, or qualified output, there is finally beginning to be an exploratory quantitative approach.

Worken is one such attempt.

Just as in the era of large language models, Tokens gradually became an important unit for measuring model invocation volume and usage cost, Worken also attempts to provide physical AI with a commercial anchor tied to actual operations, pushing industry competition to shift further from whose demo is cooler and whose concept is louder to whose robots can continuously and stably create productivity.

Therefore, the combination of Zhiyuan's new model, new body, and native OS aims to center on Worken and turn technical capabilities into effective work hours as much as possible.

Technical R&D certainly has long-term value, and not every item has to immediately become effective work hours.

But embodied intelligence ultimately has to be judged by whether it can continuously create value in real operations. And Worken is that ruler.

Hardware will depreciate, but intelligence can appreciate. It is time for the embodied intelligence field to bid farewell to the entertainment circle in the greenhouse and prove itself on real, hard-fought positions.