From Hype to Heat: The Sarvam AI Controversy Explained | FrontPage

they have a non-existent marketing channel. I didn't even know this. They need to ramp up marketing. Indic LLM is huge, I fully support sarvam. Who has the patience to fine tune an accurate indic model? I have compared whisper with Sarvam STT and Sarvam wins by a l;aarge mile and I say, the difference is HUUUGE

Sarvam did NOT make an LLM. This is a blatant lie, designed to fool the average gullible Indian.

What they have done is taken Mistral 3(small), which is the actual LLM (which only a handful of companies can create, costing over $250 Million per version to train) and they "Fine-tuned" this base model with new "indic" data. Fine-tuning means to augment the existing LLM with new data, there by generating a slightly modified LLM that performs better on a new domain (indic languages in this case).

Mistral is "open source", another misleading term here but the license allow you to distribute the modified LLM as if it's your own and run it on your own PC or cluster in a data center.

So there is no "new LLM". No one "dropped" anything amazing. Fine-tuning is a mundane process available to everyone, even with OpenAI, Llama or Gemini. A basic python programmer can pull this off.

The "indic" data is hard to get. But that's donkeys work, no real tech there. More of a data entry style job.

Pitching all this as a new LLM is disingenuous and misleading and tarnishes our standing in the international stage. Maybe Indians can help themselves...

It is a wrapper sitting on open source Mistrak. It is not a real LLM model. Unfortunately they are creating a hype. My question is if Frontier models like ChatGPT, Claude, Gemini, Llama or Grok, how can a wrapper product like Sarvam compete?

No problem, atleast gave a glitter of hope

We need our own websites and apps.

This is why India can't have it's own LLM so early. People will criticize and bring it down, if it can't compete with the top LLMs which it obviously can't.

They should stop shouting and build quietly.

Even if you create an Indian Whatsapp (hike) or insta no one will use it. There is no cure for stupidity. You can Create a billibilli like clone and people will still not use it, no one can save this country.

LLMs produce output based on the training set, at least they have collected that. In the future, they may release pre-trained models.

Step 1 or < 10 from Sarvam. We should not be ruthless. Imagine step 100 from Sarvam. There will be smart and hardworking people at R&D. They will be heartbroken if you keep criticizing them on how each step is not more significant than Google/OpenAI. Patience!

Obv. aspirations of an organization gets higher with each step.

Indian people - always complaining. Negative self talk only brings us down. This Sarvam has a different purpose than ChatGPT etc.

less downloads? I can't find it in Playstore..

Typical Indian mindset cannot appreciate who was trying and failed they think life is bollywood. In which seen was twisted and everything got better. They should watch byd series how this company grow up . After failing many times.

You have got it totally wrong. The first model of SarvamAI is actually credible and something that the Indian engineers should be proud of.
So before you judge the credibility of a model based on "downloads", let me tell you that the consumers of these models are mostly open source developer and AI community that are developing products and systems based on AI.
How many open source devs and strong community do we have in india who are working on AI ? To answer, "Very Few"!. And if they need a model, they would use llama or other open source models.
its not easy to get community popularity within India.
I remember couple of months back discussing in my office that hey "sarvam" gonna launch its on LLM. and every one in the room was like "what is sarvam?"
So please also keep this in mind , unlike a strong dev community in west which led to popularity of deepseek, here in india we have a very weak community !
Clearly, "downloads" is one of the last metric i would use to judge sarvam.
And technically speaking, dont go by number of parameters.. today its 24 Billion, tomorrow they can scale it to trillions. The point to note is they are handling Indic language which are more complex to process than "English". So sarvam is actually a hit among the hardcode open source LLM community in India and let me tell you there are only handful of devs in india who actually understand LLMs and are opinionated to use one.

WTF is this misleading thumbnail.

Don't comment anything before any expertise. People commenting like they know better than actual developers. I am not saying about speaker i am saying about people who commenting with some knowledge like a dumb. Btw you showed both perspective.

I mean i didn't even know about this llm, so how would i even download it. Thanks yo my reading habit that i saw an article on it and started researching on it. Now i will use it if it is good.

Because no one is asking for an indic LLM. Any half decent AI can interact in any language, even regional Indian ones.

Bro, if you're not good with LLMs, then don't talk, please. And "This is why India can't have its own LLM so early"—maybe, but still, we need our own real, accurate data and proper ML training processes. That is a leap—not every baby becomes a Sarva-Shastra master from birth. It has to be learned as they grow. Criticizing and all that—I don't care. But please, at least someone is doing more than nothing. So, if you can't appreciate it, then it's better not to demotivate them

Technically, they built a translator. Invested wheel again!

Have used it .Queried some.Not bad But cant download from playstore.havent found it

Let sarvam beat Gemini on answering questions in Indic languages, I’ll believe it

this is the most stupid clickbait video i have ever seen it was not a flagship model . The aim was to open source the techniques uses in pretraining are you completely out of mind moron

Too late to enter the market. There are atleast 100 models better than Sarvam.

India may start late, but just because you started the race a little late does not mean you will stay last. Let's see how Sarvam grows in future.

Satellite flop, AAA game flop, AI flop - kuch toh dhan se banao yaar. No wonder our sages gave up and went to the forests centuries ago.

No DeepSearch No WebSearch
just a model who chats
& also no ai image generation
What different they are doing ? To standout In this AI race ?

its based on Mistral..? Its not developed by them ?

The media’s fixation on sensationalizing a tweet about Sarvam AI’s Sarvam-M launch, rather than engaging with its technical substance, exposes the shallow state of tech journalism. Sarvam-M, a 24-billion-parameter hybrid language model built on Mistral Small, showcases advanced post-training techniques tailored for Indian languages, math, and programming. The accompanying technical blog details a three-step methodology: (1) Supervised Fine-Tuning (SFT) with curated datasets (30% coding/math/reasoning prompts and 50% general prompts in 10 Indian languages, with Hindi at 28% of Indic data), (2) Reinforcement Learning with Verifiable Rewards (RLVR) using a custom curriculum and GRPO algorithm for stable reward attribution, and (3) inference optimizations for efficient deployment. This resulted in significant performance gains: +20% on Indian language benchmarks, +21.6% on math, +17.6% on programming, and an impressive +86% on a romanized Indian language GSM-8K benchmark, rivaling larger models like Llama-3.3 70B. Sarvam-M’s “think mode” further enhances its reasoning capabilities for complex tasks, as demonstrated by solving JEE Advanced 2025 questions in Hindi.

The launch was a proof-of-concept for Sarvam’s data curation and training methodologies, not a flagship LLM release, yet media outlets and X commentators fixated on download numbers (334 on Hugging Face in two days) instead of the rigorous research outlined in the blog. This oversight ignores Sarvam’s contribution to India’s sovereign AI ecosystem, prioritizing clickbait over meaningful discourse on innovative AI development.

I tried their model it's no better than fine-tuning version of llama , they didn't do sft after pre training I guess it's noisy , I did not tried their paid models though.

There is nothing like building for bharat.

India can't cope 100X better models are produced daily in China and US

What is the product?

First let’s keep India clean without littering . Even African countries are much cleaner than India

Hey bros, is anyone interested in TTS pipelines for Indic languages or in taking ideas to products in the AI domain? Let me know—I'm an AI engineering student.

Have a market fit first. No one is saying its a bad innitative, it simply doesnt have a market. Who is the target audience. All people with paying capacity will prefer better quality models and would barely use indic models for that matter. Next you want to assist farmers etc, who is paying for the compute and usage there? So the reality here is. We dont need to show our progress while boasting crap. We can take our good time but atleast come up with something useful. Just because you are connected with the government system, doesnt mean we should let them have their way. Many people and organizations are doing better than them who barely see the limelight or support yet their finetuned models are far superior to what is being shown here. You can be good with indic langauges but there needs to be a market for it as well. Its not about nationalism, its about common sense. Build a product that everyone would want to use or simply build it for government and give it away for the public but then the whole purpose of running an AI businesses goes down the drain. How does one make money, who is the target audience, who will actually use it and the ROI. The reality is far beyond simply making and showing the "Indian narrative". We dont need to boost our Egos while achieving it. On the contrary what if these guys simply built services around AI for india by building infra for india. Gradually if the need comes one can show what and how to do. The same issue has been faced by OLA's offshoot as well. So you as media also need to come back to the ground and support reality. Maybe we collectively can set a better direction for the AI space in india rather than to try to satisfy our Egos fighting the west.

i try survam -m it need to imporve too moch even they not able to write the good code

In India people expects instant success on the first try, otherwise be ready to face huge criticism from people who don't even know what they are criticize about. Because of this type of environment there was not much innovation or research output. In india this phase been never been valid - " Failure is the mother of Success"

this is what happens when you celebrate too early

India does not have data. It can only build on what the Western data centres provide for free. AI initiative is farce. When for last 50 years you depends on foreigh tech and data you can't build anything than what that foreign country build. At least China had their own tech infrastructure, India just used what was offered to it.

There are plenty of issues with indic language AI not from the technical perspective but for commoditizing it. Let me list down some points:
1. Someone who does not know english (which is more or less the major language on the internet) wont really go to an app and write Bengali or Hindi or even Malayali to solve their daily problems.
2. India has a huge mobile phone userbase. Its faster to type hindi using english rather than typing hindi with a hindi keyboard. It adds more hassle than the convenience that it brings to the table.
3. I dont see a current trend in India where someone who is not very familiar with a google search could go to an Indic language llm and ask their queries or solve their problems.

The problem in a nutshell is that indic language llms are a great thing as a concept but it wont attract a large userbase or solve mission critical problems.
Thats my take on it.

Look at Veo3. That is what you call a revolutionary tech.

Gemini one day realtime translate everything. Just wait for 1 to 2 years.

The play is to survive till there is traction for acquisition by the big players. Indic data ready to be consumed.

creating own things is development of a country, but here giving license to foreign companies is associate with development. maybe i am wrong. but we are dependent on foreign companies for even basic online service like chat, we need to rethink again.

i remember when telegram launched they promoted it as its Indian origin app, and everyone was installing it. even most of the people think bata is indian company. india had hike app it was better than whatsapp but it was not promoted properly.

Anything based in India is a shitshow😂

It's definitely good ai . But it's start lagging after 20 minutes of use. And i'm having trouble finding the AI search option. 😊

Sarvam ai hire people based on sc st OBC .