When India’s most famous AI startup, Sarvam AI (Axonwise Pvt Ltd), launched its newest LLM, it prompted discussions across the firm’s strategy and broader challenges going through India’s AI mission.
One of many first firms to be chosen beneath the IndiaAI Mission to construct India’s sovereign foundational LLM, Sarvam AI not too long ago launched Sarvam-M, a 24-billion parameter open-weights hybrid language mannequin constructed on prime of Mistral Small.
A breakthrough for Indic AI researchers to construct use circumstances, the mannequin helps 10 totally different Indian languages, together with Hindi, Bengali, Gujarati, Kannada, and Malayalam, amongst others.
However the tepid response—simply 334 downloads in two days on Hugging Face—acquired some flak. Deedy Das, an investor at Menlo Ventures, known as it “embarrassing,” and stated there’s no actual viewers for this incremental work. His feedback led to a heated debate among the many Indian AI neighborhood.
Das contrasted this with an open-source mannequin, developed by two Korean faculty college students, that garnered about 200k downloads.
Whereas Sarvam claims to be dedicated to constructing a foundational mannequin, with many extra releases within the pipeline, this early launch of Sarvam-M, constructed on prime of the French AI mannequin, raised eyebrows.
Sarvam just isn’t alone. The federal government-backed BharatGen’s Param-1 mannequin, which turned out there on AIKosh final week, has solely acquired 12 downloads when writing this text.
Is there a Advantage within the Criticism?
Das’s critique goes past mannequin downloads. He argues that Sarvam’s effort is a mirrored image of misplaced ambition. “Nobody is asking for a barely higher 24B Indic mannequin. Clearly,” he stated. “If you wish to practice fashions, there ought to be an excellent cause for it.”
He additional famous that there have been cheaper and higher fashions made by Google and TWO.ai that carry out higher in all these languages. “I’ve nothing towards Sarvam, however I simply don’t assume that at this second their contributions are remotely commensurate to their funding,” he added.
In keeping with the corporate, it has raised $41 million from main buyers like Lightspeed India Companions, Peak XV Companions, Lightspeed Enterprise Companions, and Khosla Ventures, amongst others. Its valuation as of March 2025 stands at $111 million, in accordance with Tracxn.
That stated, a number of customers on X identified that the mannequin was good and had a number of use circumstances, but it surely additionally wanted enchancment.
Whereas appreciating the Bulbul TTS mannequin of Sarvam, holding some reservations, Das recommended Sarvam work on basic, ground-level rethinking of the {hardware} and software program stack, citing China’s DeepSeek for example.
“Deal with large-scale Indic and different normal mannequin knowledge assortment, which is driving the frontier fashions at this time,” he stated.
Nonetheless, in accordance with a technical report, Sarvam-M surpassed Llama-4 Scout in efficiency and holds its floor towards bigger fashions reminiscent of Llama-3.3 70B and Gemma 3 27B.
“We discovered the bottom Mistral Small mannequin could possibly be considerably improved in Indian languages,” reads the report. Nevertheless, it skilled a minor decline of 1 % in English information evaluations like MMLU.
The corporate took satisfaction within the growth. “Sarvam-M represents an necessary stepping stone on our journey to construct Sovereign AI for India,” Vivek Raghavan, the co-founder of the corporate, stated on X
Additionally, Aashay Sachdeva from Sarvam AI defended the mannequin on X, saying Sarvam-M achieved new benchmarks for Indian languages and pointed readers to the technical weblog detailing the customisation and fine-tuning course of.
Sachdeva additionally posted a Google Sheet on X, during which he requested Sarvam-M’s Assume mannequin 7 questions from JEE Superior 2025 in Hindi, and it answered all of them accurately. Clearly, the mannequin is nice for a number of use circumstances.
Conform to Disagree
Many appreciated Sarvam’s efforts and emphasised that innovation just isn’t at all times about instantaneous virality. Nevertheless, one can argue that it was all about expectation mismatch and whether or not India’s wager on homegrown AI can ship outcomes that justify the hype.
A lot of the criticism comes from evaluating Sarvam to OpenAI or DeepSeek, whereas the issue the corporate is making an attempt to resolve is basically totally different.
Harveen Singh Chadha from Sarvam AI, who was additionally essential when Krutrim launched its LLM earlier this yr, stated folks critiqued the mannequin with out really testing it.
Consumer @cneuralnetwork, who works at AI4Bharat, defended Sarvam’s work by specializing in the methodology: “The mannequin just isn’t the work—they course of how they made the mannequin is the work I beloved essentially the most. It units the bottom for different builders on methods to post-train and probably do even higher.”
Some even shared examples of individuals utilizing Sarvam’s newly launched mannequin on Google Colab, and listed use circumstances the place the mannequin might probably assist farmers and the authorized neighborhood, centred round fixing for Bharat.
In the meantime, Kurain Benoy, a machine studying engineer at Sarvam AI, echoed a broader sentiment of optimism and nationwide satisfaction. However, Gaurav Aggarwal, VP and chief AI scientist at Reliance Jio, questioned the nationalism round Sarvam, stating it’s funded by Western Buyers.
Additionally learn:Is Sarvam 2B Good Enough for Indian AI Developers?
The Unhappy State of Indic AI Analysis
India has 600 million smartphone customers, and a substantial inhabitants choose utilizing Indic keyboards. As an example, AskDisha, the AI chatbot on IRCTC web site constructed by CoRover.ai for reserving tickets in Indic languages, is a traditional case. Fashions like IndicTrans2 from AI4Bharat are constructed for the bigger inhabitants of Bharat.
The necessity to construct AI for Bharat has been highlighted a number of instances previously. A big inhabitants of Indians are additionally first-time smartphone customers, who use it of their native language as a lot as potential. This divide between English customers and Bharatiya populations presents distinctive challenges in bridging the hole between expertise and its various customers.
Many startups and researchers in India goal to resolve for Bharat and Sarvam wants to focus on the use circumstances they goal for. Pratyush Choudhury, principal at Collectively Fund, defined that most individuals outdoors India don’t perceive the challenges, and neither do they admire that compute is sort of the invisible ceiling.
“Plus, there’s a complete host of Indic-language use-cases the place this sovereign mannequin would work significantly better in comparison with utilizing every other open-weights mannequin,” he added, countering Das’ arguments.
This exhibits the significance of constructing Indic AI fashions, a development that began two years in the past when builders have been constructing on prime of Llama for languages like Kannada, Tamil, and Malayalam.
Raj Dabre, senior analysis scientist at Google and one of many creators of IndicTrans2, stated, “Earlier than Sarvam M got here out, folks have been complaining concerning the lack of IndicLLMs. After Sarvam M got here out, persons are nonetheless complaining.”
Whereas Sarvam’s work on Indic languages is much extra necessary than that of one other wrapper startup, the talk surrounding Sarvam, judging merely primarily based on downloads, signifies the event neighborhood is overlooking ‘constructing in India, for India’.
The put up Sarvam AI’s Backlash Exposes the Unhappy State of Indian AI appeared first on Analytics India Journal.
