NeuroDx 预览 Sphota：一种旨在降低内存需求的语言模型架构

NeuroDx previewed Sphota, a new language model architecture designed to represent meaning more efficiently, at Cypher 2026 in Bengaluru on October 9.

Dr. Siddharth Panwar, CEO of NeuroDx and an assistant professor at IIT Mandi, said the company expects to emerge from stealth in the coming months, with an official model release planned for later this year. The architecture is proprietary, and NeuroDx intends to keep it that way for the foreseeable future.

How Sphota works

According to Panwar, most current language models process text sequentially, building their understanding one token at a time. Sphota takes a different approach by trying to represent an event's underlying meaning rather than relying primarily on the sequence of words used to describe it.

“Intelligence is, in our opinion, not residing in language. The substrate of intelligence is somewhere else,” Panwar said. He defined intelligence as “the ability to identify higher-order semantic structures in any domain.”

The architecture is designed to maintain a compact representation of the event being modelled and update it when relevant information changes. Rather than continually expanding its representation of the conversation, Sphota aims to preserve a bounded state that captures the essential information.

A model processing the sentences token by token could lose track of the underlying event as the wording changes, Panwar argued.

“Even if the model has figured out Asha lent Ravi a book, the second sentence has to be taken very carefully,” he said, adding that there is “a possibility of losing track of that event if you are just going, token by token.”

Sphota aims to preserve what Panwar described as the invariant elements of an event.

“Something invariant in an event that has happened, something invariant in the state that you are trying to model,” he said.

If the lender changes in a subsequent sentence, the model should update its representation accordingly. Similarly, if Ravi was supposed to return the book on Tuesday but it is now Wednesday, the model should account for the uncertainty surrounding the event.

For Panwar, the objective is to represent the underlying state and its changes without repeatedly reconstructing the same information from a sequence of tokens.

Early results

NeuroDx compared Sphota with a 600 million-parameter model and reported substantially lower memory requirements for its key-value (KV) cache, which stores information used during inference. The company also said Sphota required fewer floating-point operations.

Lower memory consumption could make the architecture easier to deploy on devices such as phones and laptops, Panwar said. He also suggested the approach could reduce hallucinations, although the company has not yet published results showing that benefit.

Panwar cautioned that the figures presented at Cypher were preliminary and did not represent the model's official release. The model was built at the smallest possible scale, he said, adding that the company had better internal results.

NeuroDx said Sphota was competitive with other models of a similar size. However, it did not disclose detailed benchmark results in the presentation.

An audience member asked how the book-lending example differed from ChatGPT, which had correctly identified the underlying facts when tested with a similar prompt.

Panwar acknowledged that ChatGPT could handle the example but argued that deriving meaning from tokens takes time. He said Sphota is designed to identify the semantic structure more directly.

He also argued that Indian developers seeking to compete with Western and Chinese AI companies need to explore alternative architectures rather than follow established approaches.

“The only way to catch up with the Western and the Chinese companies is to completely move orthogonally in terms of direction, technological architecture,” he said.

The Sanskrit idea behind the name

The inspiration for Sphota came partly from Indian linguistic philosophy, Panwar said.

He contrasted approaches that build meaning sequentially from linguistic units with the ideas of Bhartrihari, an ancient Indian grammarian and philosopher associated with the concept of sphota.

Panwar described the concept as meaning that emerges as a whole rather than through the gradual assembly of individual words.

He illustrated the idea with an example involving a peacock's egg. Seeing the egg can evoke an immediate understanding of its characteristics, including how its shell might shine in different light or the sound it might make if stepped on.

“All of it immediately flashes in your brain as soon as you see the egg,” he said.

Panwar said the term sphota refers to expression or manifestation, reflecting the idea that meaning becomes apparent as a whole.

He also connected the architecture to the Sanskrit framework for the stages of speech, which describes how information is received, processed and expressed. He said Sphota operates at the level of pashyanti, a stage associated with meaning before its articulation in words.

From brain signals to language models

NeuroDx develops foundation models for electroencephalogram (EEG) and other physiological signals, with applications spanning clinical and commercial systems.

Panwar holds a PhD from IIT Delhi and degrees from Stanford University and the University of Arizona.

The company was among 12 selected under the government's IndiaAI Mission to build sovereign AI models. It had been working on a 20-billion-parameter model for EEG signal analysis before introducing MANAS-1, a 400-million-parameter brain foundation model trained on 60,000 hours of brainwave data.

Panwar said the EEG work helped shape the thinking behind Sphota. Unlike text, electrical brain signals lack natural linguistic boundaries from which to build a token dictionary.

“I tell my engineers that EEG is nothing but 20 people singing in chorus,” he said.

The team began exploring ways to identify the underlying patterns and states these signals represent. Panwar cited epilepsy as an example of a condition that can be understood through changes in the brain's state, making state-based representations relevant to the company's research.

NeuroDx is also developing a neurosemantic channel, a non-invasive headset interface designed to identify semantic structures in the brain and align them with a chosen domain.

Panwar said the technology was not intended to read thoughts. Instead, he described it as a way to provide additional context that conventional interfaces cannot easily convey.

Clarifying that the technology is not meant for thought reading, he described it instead as a high-bandwidth method for delivering additional context that sequential interfaces like typing or speaking cannot easily convey.

NeuroDx plans to share more extensive results as development progresses. Until the company releases detailed benchmarks and technical documentation, Sphota's reported memory savings and performance remain preliminary claims that will need independent evaluation.