Soket AI Wants To Turn The Language Gap In AI Into India’s Edge

Will AI models ever be as good for Indian languages as they are for English? 

Soket AI is attempting to find a real answer to this question by building large language models (LLMs) that treat Indic languages as a core, not optional add-ons.

Founded in 2019 by Abhishek Upperwal, the startup is building a sovereign open-source AI stack that serves a gamut of Indian languages. This approach has already helped the startup secure ₹177 Cr in backing under the IndiaAI Mission, including ₹162 Cr in compute resources and ₹15 Cr in non-compute support.

Soket AI’s approach differs in one important way. While multilingual AI remains a key goal, the startup is equally focused on building strong reasoning capabilities that can eventually power enterprise, government and strategic applications.

Simply put, it wants to create an indigenous AI system that can understand, reason, write code and solve complex problems in multiple languages to trim dependence on foreign AI providers.

Soket AI’s big focus is now on Project EKA. Under this, it is building a 120 Bn parameter, open source AI model that will go beyond Indian languages to address the needs of under-represented languages across the Global South, which will be launched sometime in 2027, the company claims. 

The Focus On Math, Coding And Reasoning

For many AI startups, Indian language support is the headline feature. But for Soket AI Labs, it is the starting point.

The startup’s upcoming multilingual AI model will cover more than 20 languages across the Global South, where digital representation and high-quality datasets remain limited. 

According to founder and CEO Upperwal, the idea is to crack the AI model for India first, and then apply the same approach to other underrepresented language ecosystems globally.

The logic is straightforward. If AI systems can be made to work effectively across India’s diverse languages, the same techniques can potentially be applied to other underrepresented language ecosystems worldwide.

However, Soket AI’s bigger bet is on reasoning. Which is why the startup is investing heavily in mathematics, coding and structured logic, domains that now sit at the core of frontier AI. The startup believes that strong reasoning models can become foundational infrastructure for sectors such as banking, finance, cybersecurity, defence, law and scientific research.

This reasoning philosophy for AI has shaped Soket’s work from the beginning. In February this year, the startup showed an internal prototype that let developers generate and understand code using Hindi prompts, signalling that advanced coding tools need not remain limited to English speakers.

Since then, the company has focused on refining both model design and data strategy. Rather than following existing architectures, Soket claims that it has spent nearly a year testing different approaches to find a balance between capability and efficiency. It did not disclose the exact details of its new architectural approach to AI.

For Soket AI, the real challenge is not just building a multilingual model, but building a one that is strong enough to compete on reasoning, affordable enough to scale and broad enough to matter beyond India. But it has not been an easy ride so far.

The Long Road To Soket’s Frontier AI

Many see building a frontier AI model just a matter of securing enough compute. However, for Soket, it has equally been a problem of data, research depth and specialised talent.

One of the hardest parts for the startup has been creating high-quality datasets instead of simply piling up large volumes of data. “We hit a wall where many of the content in the domain was not sufficient. Data quality was not that great,” Upperwal said. 

As public data sources fell short for advanced reasoning tasks, Soket designed its own pipelines for collecting, cleaning and structuring data, especially for Indian and other low-resource languages. 

Having solved the data problem, Soket is now parallelly also investing heavily in, what Upperwal describes as, harness engineering, or orchestration infrastructure. In practice, this means the systems that let AI run reliably in production: the layers, workflows and tooling that connect raw model intelligence to real enterprise applications. 

While foundation models tend to grab the spotlight, Soket AI’s view is that robust orchestration is just as critical for any real-world deployment. Upperwal likens his harness engineering approach to Apple’s tight integration of hardware and software. 

In Soket AI’s case, the aim is to align the foundation model layer closely with the orchestration stack so that the research work naturally translates into commercial-ready systems.

The expanding business scope has also driven the startup to grow its team. Until recently, much of the research effort was driven by the founding team itself. However, as model development, infrastructure engineering and dataset creation accelerate simultaneously, Soket AI has begun hiring specialised researchers and engineers.

This build-out is being funded through a mix of IndiaAI Mission support and new capital. To further fuel its ambitions, Upperwal said that the startup is currently looking to raise money and has already received some commitments.

While he did not share details, he noted that investor interest in frontier AI has strengthened since 2023 as VCs and PEs have grown accustomed to the long, research-heavy timelines behind foundational model work. This has enabled the startup to focus on grunt work before rolling out its products.

Soket AI’s Commercial Rollout Plan

Unlike many AI startups that prioritise rapid commercialisation, Soket AI appears willing to sacrifice speed in favour of more profound research. 

However, instead of immediately scaling to its flagship upcoming AI model, Soket AI is first building the 24 Bn-parameter model that will serve as a proof point for its architecture and data strategy. This could come by the end of 2026.

The company plans to test this smaller version with a small group of enterprise partners before scaling up to the larger model by 2027. 

The startup expects its partners to help validate the architecture, datasets and performance in real-world environments before broader deployment. The objective is not necessarily to match the latest offerings from global leaders such as Anthropic on day one. Instead, Soket AI aims to build models that address specific enterprise requirements at significantly lower costs while gradually improving through deployment feedback.

Once validated, the company plans to scale up to the larger model and expand availability. Importantly, both models will be open-sourced.

Next, to tackle unit economics, Soket AI is pursuing multiple pathways to monetise its research over time:

• Foundation Model APIs: Making the models available through APIs, allowing developers and enterprises to build applications on top of them.
• Enterprise Deployments: Working closely with enterprises to create customised AI systems and specialised models trained for specific workflows and datasets.
• Dhrit ASR Platform: Commercialising its speech recognition technology that can capture emotional cues, identify Indian entities, and automatically correct transcription errors.
• Future Audio AI Products: Expanding into text-to-speech (TTS) and other audio AI offerings currently under development.

For now, though, revenue remains secondary. “We want to do some amazing work on the research front first, and then I think the priority of business on revenue should kickstart,” the founder added.  

He also adds that Soket intends to release model weights as well as other additional assets developed during the research process. For context, model weights, which contain the AI model’s learned knowledge, allow developers to run or fine-tune the model for their own applications. 

The company has already open-sourced portions of its Indic-language data work through Indic Corpus v1, a large-scale Indic pre-training dataset.

Expanding its real-world applications, Soket AI Labs is also working with organisations such as ICRISAT (International Crops Research Institute for the Semi-Arid Tropics) to build specialised datasets. With this, the startup is helping create what it claims could become one of the largest repositories of agricultural advisory data, which could eventually train AI systems to generate farming recommendations.

All said and done, the startup’s immediate focus remains on proving that a globally competitive frontier AI model can be built from India. Whether Soket AI ultimately succeeds in matching the capabilities of the world’s leading AI labs remains an open question. 

But as India attempts to build sovereign AI infrastructure, startups like Soket AI Labs are increasingly becoming the country’s most consequential experiment. So, can Soket AI reshape the AI landscape and help India carve out its place in frontier AI? Only time will tell.

The post Soket AI Wants To Turn The Language Gap In AI Into India’s Edge appeared first on Inc42 Media.