India's largest AI startup releases Sarvam M, a model for Indian languages, but with only 23 downloads, India's domestic AI is in dire straits, but developers are grateful for the feedback



Sarvam, an AI startup from India, the world's most populous and one of the world's leading IT powers, has announced its own open weight model,

Sarvam M. However, even after two days of its release, it has barely been downloaded, sparking a debate about IndiaAI, an Indian AI initiative.

Sarvam AI's Backlash Exposes the Sad State of Indian AI
https://analyticsindiamag.com/ai-features/sarvam-ais-backlash-exposes-the-sad-state-of-indian-ai/

Sarvam, one of the first companies selected for the Indian government's sovereign AI promotion program 'IndiaAI,' released the open-weight model 'Sarvam M' based on Mistral Small on May 23, 2025. Sarvam M is optimized for Indian languages and supports 10 languages used in India, including Hindi, Bengali, Gujarati, Kannada, and Malayalam.



However, the response to Sarvam M on the AI platform Hugging Face was lukewarm, with only 23 downloads in the two days since its release .

'Sarvam, India's largest AI startup with a market cap of $1 billion, just launched their flagship LLM program. It's a 24B Mistral Small model trained on Indian data, and has been downloaded just 23 times in the two days since it was released. In contrast, two South Korean universities trained an open source model last month that was downloaded almost 200,000 times. This is shameful,' said Deedi Das of Menlo Ventures, an AI venture capital firm.

Das also expressed disappointment in India's AI efforts, saying they are not about solving important problems, but rather about 'doing cool AI like the cool AI people are doing.'



Sarvam M isn't the only Indian AI that didn't get off to a good start. BharatGen, an AI project promoted by the Indian government, released 'Param 1,' a bilingual-based model developed for English and Hindi, in May 2025, but at the time of writing, it had only been downloaded 12 times.

BharatGen - Param 1: Indic-Scale Bilingual Foundation Model
https://aikosh.indiaai.gov.in/home/models/details/bharatgen_param_1_indic_scale_bilingual_foundation_model.html



Das's comments sparked a heated debate in the Indian AI community. For example, one X (formerly Twitter) user criticized Sarvam M for its modest improvement in Indian language benchmarks (0.49 compared to 0.47 for Llama and 0.48 for Gemma), saying, 'Just because it's made in India doesn't mean we can settle for a half-baked product.'



Meanwhile, Pratyush Choudhury, an investor at the Indian venture capital firm Together Fund, argued that 'most people outside India don't know that computing is an invisible ceiling in India,' citing the fact that high-performance AI accelerators such as the H100 are not yet commercially available in India. Regarding Sarvam M, he defended, 'This is not just a vanity tweak. It is India's first open-weighted, 24B Indian language-based LLM, built amidst harsh GPU and data shortages. Judging only from HuggingFace's short-term statistical data is missing the essence.'



In addition, Alok Bishoi, who heads the Indian operations of AWL , a Japanese AI company, reviewed 64 questions that are sensitive in India, such as beef eating, the caste system, the Kashmir conflict, a territorial dispute with Pakistan and China, and the rights of sexual minorities. He pointed out that while Sarvam M has a fairly sophisticated way of thinking, it tends to lack consistency in its political views and is heavily influenced by the English-speaking urban elite.



Amid all this discussion, Indian AI developers have remained positive. Dharmesh Ba, founder of Business Hero, an Indian AI agent development company, wrote a message to Das saying, 'Good job!' after Sarvam M's popularity led to a tenfold increase in downloads to 334. Pratyush Kumar, co-founder of Sarvam, also responded to the review, saying, 'We're very happy to receive your feedback on Sarvam M. Please continue to post. We'll use it to strengthen our pipeline as we start training the sovereign model. This was particularly interesting.'

Related Posts:

in Software, Posted by log1l_ks