When India’s most noted AI startup, Sarvam AI (Axonwise Pvt Ltd), released its latest LLM, it prompted discussions around the company’s approach and broader challenges facing India’s AI mission.
One of the first companies to be selected under the IndiaAI Mission to build India’s sovereign foundational LLM, Sarvam AI recently released Sarvam-M, a 24-billion parameter open-weights hybrid language model built on top of Mistral Small.
A breakthrough for Indic AI researchers to build use cases, the model supports 10 different Indian languages, including Hindi, Bengali, Gujarati, Kannada, and Malayalam, among others.
But the tepid response—just 334 downloads in two days on Hugging Face—received some flak. Deedy Das, an investor at Menlo Ventures, called it “embarrassing,” and said there’s no real audience for this incremental work. His comments led to a heated debate among the Indian AI community.
Das contrasted this with an open-source model, developed by two Korean college students, that garnered about 200k downloads.
While Sarvam claims to be committed to building a foundational model, with many more releases in the pipeline, this early release of Sarvam-M, built on top of the French AI model, raised eyebrows.
Sarvam is not alone. The government-backed BharatGen’s Param-1 model, which became available on AIKosh last week, has only received 12 downloads when writing this article.
Is there a Merit in the Criticism?
Das’s critique goes beyond model downloads. He argues that Sarvam’s effort is a reflection of misplaced ambition. “No one is asking for a slightly better 24B Indic model. Clearly,” he said. “If you want to train models, there should be a very good reason for it.”
He further noted that there were cheaper and better models made by Google and TWO.ai that perform better in all these languages. “I have nothing against Sarvam, but I just don’t think that at this moment their contributions are remotely commensurate to their funding,” he added.
According to the company, it has raised $41 million from leading investors like Lightspeed India Partners, Peak XV Partners, Lightspeed Venture Partners, and Khosla Ventures, among others. Its valuation as of March 2025 stands at $111 million, according to Tracxn.
That said, several users on X pointed out that the model was good and had several use cases, but it also needed improvement.
While appreciating the Bulbul TTS model of Sarvam, holding some reservations, Das suggested Sarvam work on fundamental, ground-level rethinking of the hardware and software stack, citing China’s DeepSeek as an example.
“Focus on large-scale Indic and other general model data collection, which is driving the frontier models today,” he said.
Nonetheless, according to a technical report, Sarvam-M surpassed Llama-4 Scout in performance and holds its ground against larger models such as Llama-3.3 70B and Gemma 3 27B.
“We found the base Mistral Small model could be significantly improved in Indian languages,” reads the report. However, it experienced a minor decline of one percent in English knowledge evaluations like MMLU.
The company took pride in the development. “Sarvam-M represents an important stepping stone on our journey to build Sovereign AI for India,” Vivek Raghavan, the co-founder of the company, said on X
Also, Aashay Sachdeva from Sarvam AI defended the model on X, saying Sarvam-M achieved new benchmarks for Indian languages and pointed readers to the technical blog detailing the customisation and fine-tuning process.
Sachdeva also posted a Google Sheet on X, in which he asked Sarvam-M’s Think model 7 questions from JEE Advanced 2025 in Hindi, and it answered all of them correctly. Clearly, the model is good for several use cases.
Agree to Disagree
Many appreciated Sarvam’s efforts and emphasised that innovation is not always about instant virality. However, one can argue that it was all about expectation mismatch and whether India’s bet on homegrown AI can deliver results that justify the hype.
Much of the criticism comes from comparing Sarvam to OpenAI or DeepSeek, while the problem the company is trying to solve is fundamentally different.
Harveen Singh Chadha from Sarvam AI, who was also critical when Krutrim launched its LLM earlier this year, said people critiqued the model without actually testing it.
User @cneuralnetwork, who works at AI4Bharat, defended Sarvam’s work by focusing on the methodology: “The model is not the work—they process how they made the model is the work I loved the most. It sets the ground for other builders on how to post-train and possibly do even better.”
Some even shared examples of people using Sarvam’s newly released model on Google Colab, and listed use cases where the model could potentially help farmers and the legal community, centred around solving for Bharat.
Meanwhile, Kurain Benoy, a machine learning engineer at Sarvam AI, echoed a broader sentiment of optimism and national pride. On the other hand, Gaurav Aggarwal, VP and chief AI scientist at Reliance Jio, questioned the nationalism around Sarvam, pointing out it’s funded by Western Investors.
Also read: Is Sarvam 2B Good Enough for Indian AI Developers?
The Sad State of Indic AI Research
India has 600 million smartphone users, and a considerable population prefer using Indic keyboards. For instance, AskDisha, the AI chatbot on IRCTC website built by CoRover.ai for booking tickets in Indic languages, is a classic case. Models like IndicTrans2 from AI4Bharat are built for the larger population of Bharat.
The need to build AI for Bharat has been highlighted several times in the past. A large population of Indians are also first-time smartphone users, who use it in their native language as much as possible. This divide between English users and Bharatiya populations presents unique challenges in bridging the gap between technology and its diverse users.
Many startups and researchers in India aim to solve for Bharat and Sarvam needs to highlight the use cases they aim for. Pratyush Choudhury, principal at Together Fund, explained that most people outside India don’t understand the challenges, and neither do they appreciate that compute is quite the invisible ceiling.
“Plus, there’s a whole host of Indic-language use-cases where this sovereign model would work much better compared to using any other open-weights model,” he added, countering Das’ arguments.
This shows the importance of making Indic AI models, a trend that started two years ago when developers were building on top of Llama for languages like Kannada, Tamil, and Malayalam.
Raj Dabre, senior research scientist at Google and one of the creators of IndicTrans2, said, “Before Sarvam M came out, people were complaining about the lack of IndicLLMs. After Sarvam M came out, people are still complaining.”
While Sarvam’s work on Indic languages is far more important than that of another wrapper startup, the debate surrounding Sarvam, judging merely based on downloads, indicates the development community is overlooking ‘building in India, for India’.
