The patient was supposed to be resting. Instead, hooked to a 256-channel brain implant and a decoder, he was clearing levels of Black Myth: Wukong (黑神话：悟空) using nothing but intention.
No controller. No voice. Just a neural signal mapped to a cursor, then to a character on screen.
That scene, documented by NeuroXess (脑虎科技) after an April 2025 clinical trial at Shanghai Huashan Hospital, is the clearest picture yet of how far China's invasive brain-computer interface (脑机接口) work has pushed in a single year.
Key takeaways
- NeuroXess implanted a self-developed 256-channel high-throughput flexible BCI in a 19-year-old epilepsy patient and, after 19.87 hours of training, let him play Pac-Man, Battle City, Honor of Kings (王者荣耀) and Black Myth: Wukong by thought.
- The same platform reached 71.2% real-time decoding accuracy across 142 common Chinese syllables, with single-character latency under 100 ms.
- Trials ran at Fudan University's Huashan Hospital with support from the Tianqiao and Chrissy Chen Institute (TCCI, 天桥脑科学研究院) and the Chinese Academy of Sciences' Shanghai Microsystem Institute.
- NeuroXess claims to be the only invasive BCI company globally to have demonstrated both real-time motor decoding and real-time Chinese language decoding.
The games as a benchmark
The April 2025 case is striking because it used a hard, real-world target. After 19.87 hours of classic calibration paradigms — Center-out and WebGrid — the subject controlled a cursor well enough to handle both retro games and demanding modern titles. Through NeuroXess's XessOS operating system, he also browsed Bilibili and Xiaohongshu, and steered a smart wheelchair and home devices.
A useful, honest benchmark sits in the numbers: the team reported a cursor-control rate of 4.07 bits per second after those ~20 hours of training. Neuralink's first human subject, Noland, reached 4.6 bits per second after roughly 60 hours. The comparison is approximate — different tasks, different users — but it shows the performance gap between Chinese and U.S. invasive systems is far narrower than it was two years ago.
Training, not magic, produced the result. The Center-out paradigm first calibrated the model: a position-velocity Kalman filter turned 256-channel ECoG signals into cursor-velocity commands in real time. As training hours accumulated, the subject's motor-imagery control grew faster and more accurate. WebGrid then raised difficulty by forcing the cursor to specific grid cells, continuously sharpening neural control before multi-scene apps were unlocked through XessOS. That staged ramp — calibrate, then harden, then expand — is the actual clinical recipe behind the headline "plays Black Myth: Wukong."
Decoding Chinese, not just movement
Motor control is only half the story. In a separate December 2024 trial announced in January 2025, NeuroXess worked with Huashan Hospital's neurosurgery team to decode spoken Chinese in real time from a patient with a lesion in the language cortex.
Chinese is a hard target for decoding: a tonal, logographic language built from largely monosyllabic units, drawing on more dispersed brain regions than alphabetic languages. The system reached 71.2% accuracy across 142 common syllables, with sub-100 ms latency per character — described as the highest real-time Chinese decoding level reported in China to date.
For patients who have lost speech to stroke, ALS (渐冻症) or injury, that is the line between silence and communication.
Why a flexible, high-channel implant
NeuroXess was founded in 2021 and is led by chief scientist Tao Hu (陶虎), with Peng Lei (彭雷) as CEO. Its core hardware is a flexible electrode array — thin enough to reduce the inflammatory and scarring response that plagues rigid implants, and dense enough (256 channels) to record high-resolution electrical activity across brain regions.
In an earlier August 2024 case, a 21-year-old epilepsy patient received the implant and, using High Gamma band features and an LSTM decoding model with under 60 ms system latency, was playing Ping-Pong and Snake within two days of surgery. Language and motor trials could begin just 48 hours after implantation because the high channel count let surgeons map functional areas quickly.
Honest limitations
These are still small, early, and selective. Every case reported so far involved epilepsy or tumor patients whose surgery created a narrow, ethically justified window for testing — not chronic paralysis patients who would need years of stable implants. Sample sizes are single digits.
The 71.2% Chinese-syllable accuracy is real but not conversation-ready; a quarter of syllables would still be wrong in live use. And "real-time decoding" here means a research-grade pipeline, not a certified consumer or clinical product.
The comparison to Neuralink should also be read carefully: it speaks to cursor performance, not to safety profile, battery life, implant longevity, or regulatory status. NeuroXess has said a wireless version is in design and that it aims to complete Class III medical-device registration within roughly three years — a plan, not an achievement.
What readers can do now
- Researchers and clinicians should follow Huashan Hospital's BCI publications and the TCCI-supported trials for the next open datasets; language decoding is the fastest-moving subfield.
- Investors watching neurotech should separate surgical-first invasive players like NeuroXess from non-invasive consumer BCI firms — the regulatory and reimbursement paths are completely different.
- Anyone curious about the field can try a non-invasive EEG headset to feel how noisy real brain signals are; it builds healthy skepticism about "mind control" headlines.