A humanoid robot that can dance has entertainment value. A humanoid robot that can hold a conversation — knowing who is speaking, when to interject, and when to stay quiet in a noisy theme park — has business value. Agibot (智元机器人), the Shanghai robotics company co-founded by former Huawei executive Deng Taihua and ex-"Genius Youth" engineer Peng Zhihui, has spent 2026 building exactly that layer, and its interaction model WITA-Omni (硅光动语) has now crossed a regulatory line no domestic rival has crossed before it.
The filing that matters
In May, Agibot released WITA-Omni, a robot-native end-to-end full-modal interaction model, at its partner conference in Hong Kong. Days later, on May 18, the model completed compliance filing with the Shanghai Cyberspace Administration, and Chinese media reports described it as the first embodied interaction model in China to complete large-model filing (备案) — the registration step required before generative AI services can be offered to the Chinese public. That filing does not certify quality, but it does clear the path for commercial deployment in public-facing settings: theme parks, retail floors, guided tours.
The model's architecture is built for conversation rather than question-answering. A three-layer design — Thinker, Talker, Actor — runs in parallel: the Thinker maps vision, audio, and text into one semantic space for joint reasoning; the Talker streams natural speech; the Actor synchronously generates motion and facial expressions. Because all three share the same timeline, interaction latency is compressed to within roughly 500 milliseconds, and the robot can be interrupted, corrected, and redirected mid-sentence. It supports more than 100 languages and dialects without mode switching.
Benchmarks and the flagship upgrade
The preview version of WITA-Omni topped the DailyOmni full-modal understanding benchmark, compiled by a Fudan University research team, with a composite score of 85.21 in late July — first overall and in six of eight sub-metrics, ahead of entries from Google, Alibaba's Qwen (通义千问), and ByteDance's Doubao (豆包). DailyOmni tests something robots specifically need: judging who is speaking, ordering events in time, and deciding whether a response is warranted in noisy, multi-person scenes.
On October 29 at Agibot's Innovation Day in Shanghai, the company shipped WITA Omni 1.0, the production version. At the event, an Agibot Expedition A3-Ultra humanoid held a ten-minute live interaction running entirely on the robot's onboard compute — the company says it is the first large-parameter interaction model deployed on a mass-production robot body, a design choice that addresses privacy, offline operation, and real-time latency at once. The company reports the model now leads four benchmarks including DailyOmni, the video temporal grounding list ActivityNet-TimeLens, the audio understanding benchmark MMAU, and the embodied spatial understanding benchmark EmbSpatialBench. As always, leaderboard positions reflect the benchmark suites' own definitions of ability, not field performance.
A week of commercial signals
Alongside the model story, brokerage research circulated this week summarizing Agibot's deployment momentum — and the numbers sketch an unusually concrete commercial picture for the humanoid sector:
- Scale. The company's 10,000th general-purpose embodied robot rolled off the line in March 2026, a milestone documented by state media.
- Industrial orders. In October 2025, Agibot won a cooperation deal with Longcheer Technology (龙旗科技), a global ODM for smart devices, reported at the time as one of the largest industrial embodied-robot orders in China; its Elf G2 robots now feed tablets onto test stations on Longcheer's production line. At auto-parts supplier Joyson Electronics' (均胜电子) Ningbo factory, G2 units perform flexible assembly of tripod joints with a fastest cycle time of 12.97 seconds and a reported success rate above 99%.
- The supply chain plays too. Lingjiedian (临界点), the dexterous-hand company spun out of Agibot in January, has raised multiple rounds this year at a valuation above US$1 billion, with cumulative production of over 15,000 dexterous hands this year as of October.
The interaction model is the third leg of Agibot's stated "three intelligences" architecture — movement, interaction, and task execution — and the filing plus benchmark results suggest the company sees conversation as the differentiator for service-sector deployment, where Lingxi-series robots already work as store greeters and rental performers.
Honest limitations
"First to complete filing" is a regulatory and marketing distinction, not a quality ranking — it says the model cleared the process, not that it outperforms rivals in deployment. The benchmark victories are computed on evaluation suites, and the DailyOmni ranking compares models under that team's task definitions; independent field studies of interaction quality in crowded, real venues are not yet available. The industrial numbers above are company and brokerage disclosures: the Joyson success rate is the operator's own figure from a specific station, and "one of the largest orders" for Longcheer was reported without an audited contract value. Lingjiedian's valuation is private-market pricing, which has its own cycle. Finally, filing status applies to the Chinese regulatory context; overseas deployment of the same model would face separate regimes.
What readers can do now
- Watch latency claims against live demos, not slides. The sub-500-millisecond interaction figure is the single most user-visible spec; video from the Innovation Day ten-minute interaction lets you judge continuity and interruption handling yourself.
- Separate the three business lines. Agibot's story mixes model benchmarks, robot shipments, and component spin-outs; tracking them separately — filing, units deployed, external orders — gives a cleaner read than any single number.
- If you operate public-facing venues, file status is a practical filter. Any conversational robot deployed to the public in China needs a filed model; asking vendors for their filing status is a legitimate procurement checkpoint.
- Track the second data point on Joyson's line. One station at 99% is a pilot; a second factory, a second task type, or a disclosed downtime rate would turn it into a trend.
