A 35-year-old Shanghai woman with epilepsy was fitted with a Chinese brain implant, and the device decoded her thoughts in Mandarin at 300 characters a minute, faster than a native Mandarin speaker can talk. The trial was disclosed in Nature on 19 May 2026 by Tiger Tao, co-founder and chief scientist of Shanghai BCI startup NeuroXess. The figure sits well above the roughly 220 characters per minute that a native Mandarin first-language speaker averages, and represents the first time a Chinese BCI company has publicly claimed real-time speech decoding faster than the patient could actually speak. NeuroXess has not yet published the paper. Tao told Nature the underlying papers are being written now.

What is new vs what is already peer-reviewed

The 300 character per minute claim sits on top of a separate, already peer-reviewed result from the same company. On 5 November 2025, Science Advances published “Real-time decoding of full-spectrum Chinese using brain-computer interface” (Wu et al., DOI 10.1126/sciadv.adz9968), a paper co-authored by Jinsong Wu of Fudan University and several NeuroXess scientists. That paper reported a 256-channel microelectrocorticographic (µECoG) BCI decoding 394 distinct Mandarin syllables at 71.2% median accuracy in a 43-year-old woman with epilepsy. The system was reported as running about five times slower than typical Chinese conversational speech.

The trial Tao described to Nature is a different patient (35-year-old woman with epilepsy) and a different system layer: a large language model sitting on top of the cortical signal decoder, generating Mandarin output in real time at 300 cpm. The peer-reviewed result is the syllable identification accuracy baseline. The Nature-disclosed result is the throughput claim with an LLM in the loop. Both are anchored to the same 256-channel surface-cortex implant architecture, with the same clinical site at Huashan Hospital in Shanghai.

The October cursor control trial

Nature also disclosed an October trial in which a 28-year-old man with a spinal cord injury was fitted with the NeuroXess implant and used it to control household appliances by moving a computer cursor with his thoughts, turning devices on and off through an app. The October cursor result extends NeuroXess’s clinical population from epilepsy patients to spinal cord injury patients, which is the more commercially relevant indication for the company’s stated path to chronic home use.

The implant is placed in a shallow recess in the skull, with sensors fitted on the cerebral cortex. A wire runs from the skull module to a data transmitter that doubles as a battery, embedded in the recipient’s chest. Nature appended a clarification on 20 May 2026 specifically refining the position of the NeuroXess implant relative to the cerebral cortex. Tao has previously framed the company’s approach as avoiding the brain-tissue damage associated with penetrating electrode arrays of the kind Neuralink uses.

How this fits the broader Chinese commercial inflection

Three Chinese commercial-stage entities now hold meaningful clinical or regulatory milestones in implantable BCI. Neuracle’s NEO received the world’s first commercial market approval from China’s NMPA in March 2026 and is moving toward a STAR Market listing. The NeuCyber / Beinao-1 brain-spine programme at Xuanwu Hospital has disclosed one-year follow-up in a single combined brain-spine implant patient. NeuroXess is the third entity in this group, and the only one publicly disclosing an LLM-decoder layer running on a clinical patient.

For a Fortune 500 strategist tracking the BCI category, the structural read is that Chinese commercialisation is no longer trailing the US on the AI-decoder layer. The Western narrative through 2024 and 2025 was that China was ahead on hardware approvals and manufacturing scale, with the speech-decoder and cursor-decoder LLM work still being built by Western firms (Synchron, Neuralink, Paradromics, Precision Neuroscience). The Nature piece flips the second half of that sentence. NeuroXess is talking about the LLM-decoder result as the headline clinical signal, with the hardware as established underlying infrastructure.

The state policy backdrop

Chinese state strategy has explicitly named BCI as a target sector. The government has stated it wants the country to be a global BCI leader by the end of the decade, with technical breakthroughs by 2027 and two or three world-class firms by 2030. Beijing released ethical guidelines for BCI clinical work in 2024, requiring written consent from trial participants or their guardians and ethics review of each trial. Tao framed the regulatory environment as constructive in his Nature interview, with the government proactive in issuing guidelines that companies and researchers must follow.

A separate strand in the Nature piece, surfaced by Meicen Sun of the University of Illinois Urbana-Champaign, argues that Chinese consumers’ relative comfort with corporate data access creates a self-reinforcing data loop: Chinese BCI firms get more training data, the data improves their models, the better models build user confidence, more users supply more data. The argument is contested but worth tracking as a long-arc competitive variable that the Western neural-data privacy framework explicitly attempts to constrain.

What to watch

The first signal is whether NeuroXess publishes the underlying paper for the 300 character per minute Mandarin LLM result. Tao told Nature the paper is being written. A peer-reviewed publication or preprint moves the figure from company-asserted to industry-cited. The second signal is whether NeuroXess discloses additional patients on the LLM architecture with non-epilepsy indications, particularly ALS or locked-in syndrome, which is the patient cohort where Western speech BCIs have anchored. The current Mandarin throughput result remains a single epilepsy patient and reads as a proof point rather than a cohort. The third signal is the wireless and battery-free transition NeuroXess flagged in early 2025, which would remove the current external-equipment requirement and align the architecture closer to commercial deployment standards.

Sources

• China moves AI brain implants from trials towards real-world use (Nature, 19 May 2026, clarification 20 May 2026)
• Real-time decoding of full-spectrum Chinese using brain-computer interface (Science Advances, 5 November 2025, DOI 10.1126/sciadv.adz9968)
• BCI technology may give voice to Mandarin speakers with paralysis (ALS News Today, 13 November 2025)
• In Shanghai, Brain Implants Turn Thought Into Speech and Movement (Sixth Tone, 6 January 2025)
• NeuroXess Announces Results of Combined Motor and Language Decoding Technology (Newsfile Corp / Yahoo Finance, 10 January 2025)
• Tiger H. Tao at Tianqiao & Chrissy Chen Institute