India’s first open-sourced AI model
While all the hustle and bustle between DeepSeek, OpenAI, Google, Mistral and Anthropic, India has finally broken the shackles and has released their first open-source AI model, and it’s quite impressive as well, i.e. Sarvam-M.
Data Science in Your Pocket – No Rocket Science
What is Sarvam M?
Sarvam M is a 24-billion parameter open-source AI model developed by Bengaluru-based startup Sarvam AI. It’s trained to be multilingual, mathematically smart, and reasoning-capable — all tailored for the Indian context.
In plain English? It can chat with you in Hindi, solve math word problems in Tamil, summarise news in Bengali, and even write Python code — all while understanding the nuances of how Indians speak, write, and think.
Why is Sarvam-M a Big Deal?
1. Designed for Indian Languages
Unlike most LLMs that treat non-English languages like an afterthought, Sarvam M was built from the ground up to support over 10 major Indian languages, including Hindi, Bengali, Gujarati, Kannada, and Malayalam.
It uses a tokeniser that’s super-efficient for Indian scripts. Most models bloat Indian text into 3+ tokens per word — Sarvam M keeps it lean at around 1.4 to 2.1 tokens per word. That’s smoother, faster, and cheaper to run.
2. Supercharged Reasoning and Math Skills
Sarvam M isn’t just a multilingual poet — it’s got brains. Thanks to a blend of supervised training and reinforcement learning with verifiable rewards (RLVR), it excels at:
- Math word problems
- Step-by-step logical reasoning
- Code generation and debugging
Think of it as a model that doesn’t just guess — it checks its own work. That’s what makes it useful for actual problem-solving, not just casual chat.
3. Efficient and Optimised
Even with its 24B parameter size, Sarvam M runs faster than you’d expect. That’s because it’s built on a Mistral-style transformer architecture — a sleek, decoder-only design with smart upgrades like Sliding Window Attention (SWA) and Rotary Positional Embeddings (RoPE).
In simple terms: it handles long documents without frying your GPU and still gives accurate, context-aware outputs.
BenchMarks and metrics
Summarising it all,
- Task & Conversational: Handles instructions and dialogue smoothly; performs like a solid assistant.
- Indic Language: Excels in Indian languages — top-tier for translation, generation, and comprehension.
- Programming: Strong at code generation and problem-solving; reliable for dev tasks.
- Math: Good at reasoning and step-by-step math problems; performs consistently.
- General Knowledge: Well-rounded, especially strong in Indian-context reasoning and trivia.
Real-World Use Cases
So, where does Sarvam M shine in the real world? Here are a few areas:
- Education: Multilingual tutoring platforms and test-prep bots that speak the student’s native language.
- Customer Service: Voice-based agents for WhatsApp or phone support — imagine solving problems in Hinglish or Kannada.
- Legal and Compliance: Drafting regulatory docs or answering policy queries in regional languages.
- Agriculture: AI assistants guiding farmers with local dialect support — weather, crops, loans, you name it.
Challenges and The Road Ahead
While Sarvam M is impressive, it’s still early days. Adoption has been slow — downloads on Hugging Face were modest in the initial days. But that’s expected. What matters more is the foundation it’s laying.
With backing from India’s Digital India and IndiaAI missions — and partnerships with players like Microsoft — Sarvam AI is in a great position to scale and build more fine-tuned models, open tools, and industry use cases.
How to use Sarvam M?
Being open-sourced model weights are available on Hugging Face.
sarvamai/sarvam-m · Hugging Face
If you don’t want to use it in your local system, it can also be tried out in the Playground at the link below.
Final Thoughts
Sarvam M isn’t just a model — it’s a movement.
It marks India’s arrival on the global LLM scene, showing that high-quality, multilingual, open-source AI doesn’t have to come from Silicon Valley alone. With smart design choices, India-centric training, and a focus on real-world impact, Sarvam M sets the stage for a more inclusive AI future.
Whether you’re an AI researcher, product builder, or just someone excited about the next big thing — keep your eyes on Sarvam-M. It speaks your language, literally.
Sarvam-M: India’s DeepSeek is here was originally published in Data Science in Your Pocket on Medium, where people are continuing the conversation by highlighting and responding to this story.
