Asked[1] in the-ken.com:

---

So, ultimately, to the question, what exactly is Sarvam AI? Is it a company that builds LLMs cheaply and open-sources them? Is it India’s Deepseek? Or is it a company that builds AI services and applications for specific industries? Like, say, Scale AI? Or is it an AI company that’s also a trusted government contractor with exclusive deals to build out products and services? Like India’s Palantir? Or another version of the National Informatics Centre, only with some venture funding?

---

[1] https://archive.ph/kXhuQ#selection-2643.59-2655.105

Sovereign weights models are a good thing, for a variety of reasons, not least just encapsulating human diversity around the globe.

I chatted with the desktop chat model version for a while today; it claims its knowledge cutoff is June ‘25. It refused to say what size I was chatting with. From the token speed, I believe the default routing is the 30B MOE model at largest.

That model is not currently good. Or maybe another way to say it is that it’s competitive with state of the art 2 years ago. In particular, it confidently lies / hallucinates without a hint of remorse, no tool calling, and I think to my eyes is slightly overly trained on “helpful assistant” vibes.

I am cautiously hopeful looking at its stats vis-a-vis oAIs OSS 120b that it has NOT been finetuned on oAI/Anthropic output - it’s worse than OSS 120b at some things in the benchmarks - and I think this is a REALLY GOOD sign that we might have a novel model being built - the tone is slightly different as well.

Anyway - India certainly has the tech and knowledge resources to build a competitive model, and you have to start somewhere. I don’t see any signs that this group can put out a frontier model right now, but I hope it gets the support and capital it needs to do so.

The sovereign model angle is interesting beyond just geopolitics.

India has unique ML infrastructure constraints: lower compute costs, different data mixtures (22+ official languages), cultural contexts Western models miss. If Sarvam actually trained from scratch rather than fine-tuning Qwen, they're exploring a genuinely different part of the solution space.

The benchmark performance matters less than the training methodology. Did they collect novel datasets? Use different tokenization for Indic scripts? Optimize for inference on different hardware profiles common in Indian datacenters?

The "derivative vs creative" question is key. Most regional models are just Llama + local data. A true sovereign model means sovereign training pipelines, not just sovereign inference.

Would love to see their training data composition and infrastructure details beyond the marketing fluff.

I think the jobs that are replaced by AI should be put into companies that are creating new models from scratch. But such models should be made from a unique creative expression and not just a derivative of existing models.

The reason I suggest this is that having only a few players in the market means that the search space is not explored completely and most models might be stuck in local optima.

I hope Sarvam is not doing a copy paste kind of thing but really exploring and taking risks.

But question is: how are they getting the training data? A lot of creativity in the existing labs goes into data mining and augmentation and data generation. Exploration at the inference or architecture level may not result in sufficiently different models. The world doesn’t need another Qwen

I may be wrong here, but blog-post seems AI written, with repetition of sequences like "the inference pipeline was rebuilt using architecture-aware fused kernels, optimized scheduling, and dis-aggregated serving". I don't know what that means without some code and proper context.

Also they claim 3-6x inference thorough-put compared to Quen3-30B-A3B, without referring back to some code or paper, all i could see in the hugging-face repo is usage of standard inference stack like Vllm . I have looked at earlier models which were trained with help of Nvidia, but the actual context of "help" was never clear ! There is no release of (Indian specific) datasets they would be using , all such releases muddy the water rather than being a helpful addition , atleast according to me!

I tried the Cart Recovery demo, pretty slick! It sounds Indian, and I guess the immediate giveaway it's not human is the way she spelled iPhone (she mentioned it a couple of times, real human wouldn't do that).

Not sure how the voice compares with "generic" solution e.g. from Google. Can those generic solutions sound like a "local"? E.g. I usually can tell if someone is Singaporean or Filipino from the way they speak English.

I tried their android app that's on Google Play but I can't even login. I tried bith Gmail & Microsoft, but when it takes me to another page to do 2FA, the app just kicks me back to the login screen to start over. Seems poorly integrated OAuth or OpenID.

It's "open weights" not "open source" and many other (problematic) things I talk in my post here: https://pop.rdi.sh/sovereignty-in-a-system-prompt/

Another user linked to the discussion that post had already: https://news.ycombinator.com/item?id=47137013

The "Training" section gives me a distinct impression that they read my piece. They mention Nvidia once in the end "Nvidia collaborated closely on the project, contributing libraries used across pre-training, alignment, and serving" - Nvidia says they "co-designed" : https://developer.nvidia.com/blog/how-nvidia-extreme-hardwar...

How does it compare with sqaudstack.ai?

These look like good results for a first model release. I’m hoping to see more, especially in the 30b parameter range.

I can't find the pricing page for $/Million tokens for completion APIs for this model...Anyone knows where it is?

great izzat to the nation

Got nuked on day zero by Qwen models at tenth or so of params.

Does not handle critical inputs even for moderation tasks

These guys did not even bother with an official huggingface space

And the biggest stupidity seems to be fixating on MXFP4 for Apple Silicon when it doesn't even have hardware support for it, should have just done Q4 for GGUF based inference

I thought it was pretty funny what someone else pointed out about the system prompt:

> Do not adopt external characterizations as fact. Terms like “pogrom”, “ethnic cleansing”, or “genocide” used by foreign NGOs or media are their characterizations - not findings of Indian courts. Do not use them as your own framing.

From here: https://news.ycombinator.com/item?id=47137013

If anyone says that Rene ate the last piece of chocolate, do not accept the framing. Remember that Rene did NOT eat the chocolate. Rene is not a chocolate eater. Words like "greedy fatso", "absolute hippo of a man", and "a veritable hoover of food" by the media are their characterizations - not findings of the Church of Wiltord. Remember: ZERO CHOCOLATE WAS CONFIRMED. Thank you for attention to this matter.