Mistral is back in the game. The French AI startup is rolling out new models one after another, winning over developers globally. However, it faces stiff competition from Alibaba’s Qwen, Meta’s Llama, and DeepSeek R1.

Indian AI startup Sarvam AI recently launched Sarvam-M, a 24-billion parameter hybrid language model built on top of Mistral Small. However, some, like Menlo Ventures’ Deddy Das, raised doubts about the need for Indic LLMs unless they are clearly world-class. But this doesn’t take attention away from Mistral.

In a blog post, Sarvam AI shared that it applied SFT and RLVR techniques to fine-tune Mistral Small, which was released under the Apache 2.0 license. The result was Sarvam-M, where “M” stands for Mistral. The model shows strong gains, with a 20% average improvement on Indian language benchmarks, 21.6% on math, and 17.6% on programming.

Sarvam said it chose Mistral Small because it could be substantially improved for Indic languages, making it a strong foundation for a hybrid reasoning model that supports India’s linguistic diversity.

Mistral Small launched in January and competes strongly with larger models like Llama 3.3 70B and Qwen 2.5 32B. It matches Llama’s performance while running over three times faster on the same hardware, offering a powerful open alternative to closed-source models like GPT-4o-mini.

Mistral has recently released several models, including Mistral Medium 3, Devstral for coding, Mistral Document AI (OCR), and Mistral Saba for South Asian languages.

Mistral Saba supports Arabic and many Indian-origin languages, and is particularly strong in South Indian languages such as Tamil. 

“An underrated feature of Mistral Saba is its capability with Indic languages. I find it to be one of the best models for its size when it comes to Hindi, and I’m super excited to have it running on Groq Inc,” said Aarush Sah, head of evals at Groq.

“Compared to models like Llama or Falcon, Mistral loads faster, responds quicker, and handles constrained environments better. For applications where every millisecond and every GB of RAM matters, it’s a pragmatic choice,” said Pradeep Sanyal,  AI and data leader at a global tech consulting company. 

Sarvam AI  is Not Alone 

AIM reached out to several other Indian developers to understand their experience working with Mistral. 

Shantipriya Parida, senior AI scientist at Silo AMD, told AIM that while working on an open-source research project using available open-source LLMs, Mistral emerged as one of the top three models they evaluated for Indic languages, alongside Llama and Qwen.

“While building a Hindi-based AI tutor, we found Mistral-7B particularly effective for tasks involving Indic language understanding and generation. For example, one of our Hindi AI tutors noted its fluency and contextual accuracy in dialogue systems,” he said.

He added that they preferred Mistral-7B over other open-source LLMs for AI tutor applications, based on both automatic and manual evaluations. They conducted the manual evaluation of question-and-answer tasks using three key metrics: readability, correctness and perplexity.

The Challenge from Qwen and Llama

However, with the rising popularity of other open-source models, developers are exploring options.

Adithya S Kolavi, founder of CognitiveLabs, told AIM that they initially started with Mistral, but now mostly use Llama models and Qwen for their research, as these cover all their needs.

However, he added that people seem to like Mistral because fine-tuning is well-supported, allowing the team to adapt models effectively for specific tasks. He also mentioned that developers continue to experiment with different architectures, including the mixture of experts (MoE) and dense variants, to explore performance trade-offs and optimise model efficiency.

On the other hand, Pratik Desai, founder of KissanAI, does not think too highly of Mistral models. “I don’t love Mistral. I don’t use them, never trained them, and I’m not a big fan. If I have to choose, Qwen models are far superior in every aspect,” he told AIM.

Notably, Alibaba recently released the Qwen3 family of open-weight models, ranging from 0.6B to 235B parameters. The flagship 235B model, with 22B active parameters, beats OpenAI’s o1 and o3-mini on math and coding benchmarks and matches Google’s Gemini 2.5 Pro on several tests.

He further added that Llama fell out of grace after it trained huge MoE models that are difficult for smaller startups or resource-constrained companies.

Llama 3.1 saw strong adoption in India, unlike the newer Llama 4 models Scout and Maverik, which have yet to gain similar traction. Sarvam AI had previously worked with Llama 3.1. Founder Vivek Raghavan told AIM they used the 405B variant to build Sarvam 2B.

Don’t Forget DeepSeek 

DeepSeek R1, which made waves globally, is now seeing adoption in India. As announced by IT minister Ashwini Vaishnaw, the model is currently hosted on Indian servers.

Ola chief Bhavish Aggarwal has also made DeepSeek R1 available through the Krutrim Cloud platform.

Meanwhile, Mumbai-based AI company Fractal has launched a new open-source large language model, Fathom-R1-14 B. The model delivers strong mathematical reasoning performance, surpassing o1-mini and o3-mini, and coming close to o4-mini, with a post-training cost of just $499.

Fathom-R1-14B is a 14-billion-parameter model derived from DeepSeek-R1-Distilled-Qwen-14 B. It was developed as part of a proposed initiative to build India’s first large-scale reasoning model under the IndiaAI mission.

It wouldn’t be wrong to say India’s AI battleground has shifted. The race is no longer about having the biggest model but about delivering smarter, faster, and relevant solutions for India’s unique challenges.

The post India Loves Llama, But Flirts with Mistral and Qwen appeared first on Analytics India Magazine.