Why Sarvam AI's launch is stirring up controversy and concern
26 May 2025

Sarvam AI, a leading player in India's artificial intelligence (AI) space, has launched its latest language model, Sarvam-M.

The model marks a major milestone for Indic AI researchers and supports 10 different Indian languages, including Hindi, Bengali, Gujarati, Kannada, and Malayalam.

However, the initial response to the release was rather underwhelming, with less than 720 downloads on Hugging Face within three days.

Sarvam-M is a 24-billion multilingual, hybrid-reasoning, text-only language system built on top of French AI model Mistral Small.

Investor criticizes Sarvam-M's performance
Investor feedback

Deedy Das, an investor at Menlo Ventures, slammed Sarvam-M's performance, calling it "embarrassing" and questioning its relevance.

He argued that there isn't a significant audience for this incremental work.

Das compared it to an open-source model created by two Korean college students, which got some 200,000 downloads.

He also pointed out that more cost-effective and higher-performing models already exist.

Despite the criticism, Sarvam AI stands by its commitment to develop a foundational model and has many more releases planned.

No one's asking for a marginally better 24B Indic model
Scenario

Das's criticism extends well beyond just downloading models. "No one is asking for a slightly better 24B Indic model. Clearly. If you want to train models, there should be a very good reason for it," he said.

"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.

Sarvam has raised $41 million from top investors. As of March, its valuation stood at $111 million, per Tracxn.

BharatGen's Param-1 model also struggles to gain traction
Performance comparison

Sarvam is not alone. The government-backed BharatGen's Param-1 model, which was released on AIKosh last week, has only seen a paltry 12 downloads.

This further emphasizes the difficulties Indian AI start-ups have been facing in getting their products to stick.

Despite these issues, Sarvam AI's co-founder, Vivek Raghavan, is hopeful about their new language model.

He said that "Sarvam-M represents an important stepping stone on our journey to build Sovereign AI for India."

Sarvam-M outperformed Meta's Llama-4 Scout in performance
Defense

In defense of its model, Sarvam AI shared a technical report that showed Sarvam-M outperformed Meta's Llama-4 Scout in performance and held its own against larger models such as Llama-3.3 70B and Google's Gemma 3 27B.

The report noted a minor decline of 1% in English knowledge evaluations like MMLU but emphasized significant improvement in Indian languages.

Aashay Sachdeva from Sarvam AI defended the model on X, stating that it achieved new benchmarks for Indian languages.

Sarvam-M's potential use cases highlighted
Use cases

Sachdeva also shared a Google Sheet showcasing Sarvam-M's capabilities, where it answered all seven questions from JEE Advanced 2025 in Hindi correctly. This indicates the model can be useful for various applications.

Despite the criticism, many users on X have noted that while the model is good and has several use cases, it also needs improvement.

Some even shared examples of people using Sarvam's newly released model on Google Colab and listed potential applications in agriculture and law.