Was the AI translation live?
What is finetuning? If it is inspecting and making changes to the entire Qwen 3 30B A3B model then doesn’t it sound like double the work?
Basically the models you see and use are called Instruct (or Reasoning Models) which are trained specifically to work as assistants.
Base Models are raw models that are trained to create Instruct Models.
You can potentially take a Base Model and train it to create your own bespoke Instruct Model.
Then you can train Instruct Models to do even more stuff.
Models in general are trained multiple times for different use cases.
Have a friend who works at Sarvam, can confirm that these are not fine tunes.
Don’t know about the architecture part that was mentioned above.
Thanks!
You have said it yourself. This isn’t a new trend in the software engineering industry. And by software engineering industry I mean globally.
They wont let me try it out without logging in. However, since “India” is mentioned 12 times on their front page, I am very happy.
WTF is wrong with these companies. When I go to perplexity.ai or gemini.google.com, they don’t force me to hand over my info or read their marketing spiel before letting me try out their software. I can use full capability of translate.google.co.in without logging in.
I was looking for a subscription after my perplexity pro expired. It prob wont be sarvam.
1 Like
I have been using their sarvam-m model (through API) for the last one week. It is impressive, and can potentially replace Qwen 80B for my use case.
What’s your use-case with Qwen-80B?
Open Source models are available here: sarvamai (Sarvam AI)
Looking forward to how this turns out in the end
RAG query of Sanskrit/Hindi/English text
1 Like
These are all old models. The new ones have not been added to huggingface yet from what I can see.
Has anyone compared responses in natural indic languages? I would love to see a comparison with Gemini, which is exceptionally proficient at it. This applies even to the free version of Gemini too.
I showed what “an AI” can do to an old family member this morning. Half an hour later they were making poems and composing songs on Gemini. No typing.
Oh wow. I hope it helps bridge the divide between english and other languages here.
This is actually pretty decent
I stand corrected.
Looks like standard but entirely new Architecture.
GGUF Llama support soon: Feature Request: Support for Sarvam-30b and Sarvam-105b · Issue #20175 · ggml-org/llama.cpp · GitHub
Unfortunately, since it’s not multimodal, it doesn’t interest me much.