Anna Upreti retweeted
We're open-sourcing two frameworks for evaluating Indian ASR, and a full guide on evaluation across 22 languages.
WER (Word Error Rate) and CER (Character Error Rate) were built for languages like English. Indian languages behave differently. Formal and colloquial forms coexist. English words move fluidly between scripts.
In this guide, we introduce four complementary metrics LLM-WER, LLM-CER, Intent Score, and Entity Preservation Score. In combination they capture a more complete view of Indian ASR system performance.
Read more: sarvam.ai/blogs/evaluati…
Both frameworks are available to use and build on for evaluating Indian ASR in real-world settings.
Open-source frameworks:
LLM-WER/CER: github.com/sarvamai/llm_w…
Intent + Entity Score: github.com/sarvamai/llm_i…
WER (Word Error Rate) and CER (Character Error Rate) were built for languages like English. Indian languages behave differently. Formal and colloquial forms coexist. English words move fluidly between scripts.
In this guide, we introduce four complementary metrics LLM-WER, LLM-CER, Intent Score, and Entity Preservation Score. In combination they capture a more complete view of Indian ASR system performance.
Read more: sarvam.ai/blogs/evaluati…
Both frameworks are available to use and build on for evaluating Indian ASR in real-world settings.
Open-source frameworks:
LLM-WER/CER: github.com/sarvamai/llm_w…
Intent + Entity Score: github.com/sarvamai/llm_i…
7
62
515
29,322
It’s finally out!
A lot of hard work went into this over the last 10 months, excited to see what people build with it.
Try it out here and let us know what you think:
docs.sarvam.ai/api-reference-…
A lot of hard work went into this over the last 10 months, excited to see what people build with it.
Try it out here and let us know what you think:
docs.sarvam.ai/api-reference-…
📢 Open-sourcing the Sarvam 30B and 105B models! Trained from scratch with all data, model research and inference optimisation done in-house, these models punch above their weight in most global benchmarks plus excel in Indian languages.
Get the weights at Hugging Face and AIKosh. Thanks to the good folks at SGLang for day 0 support, vLLM support coming soon. Links, benchmark scores, examples, and more in our blog - sarvam.ai/blogs/sarvam-3…
4
4
87
3,955
Anna Upreti retweeted
Btw have been test running @SarvamAI this morning is Malayalam. Voice mode. All I can say is.
If this had come out when mom was around, with all her interest in research & culture, she probably would’ve had a much better time in her final days.
If this had come out when mom was around, with all her interest in research & culture, she probably would’ve had a much better time in her final days.
3
17
205
5,331
Anna Upreti retweeted
Hard to believe Sarvam AI able to beat o3-mini and Deepseek R1
but ya it's true
but ya it's true
🚨 o3-mini crushed DeepSeek R1 🚨
"write a Python program that shows a ball bouncing inside a spinning hexagon. The ball should be affected by gravity and friction, and it must bounce off the rotating walls realistically"
6
85
743
39,802
Anna Upreti retweeted
Models team
@_mohit_singla @anand_404 @kediaharshit9 @rkal4 @AashaySachdeva @sumanthd17 @ArpitDwivedi100 @sushil_khyalia @ManavSinghal157 @sohampetkar
Missing @selfawareatom @AnnaUpreti anand
@_mohit_singla @anand_404 @kediaharshit9 @rkal4 @AashaySachdeva @sumanthd17 @ArpitDwivedi100 @sushil_khyalia @ManavSinghal157 @sohampetkar
Missing @selfawareatom @AnnaUpreti anand
90
324
3,104
61,518
Anna Upreti retweeted
Introducing Sarvam Akshar: a workbench for solving last-mile problems in knowledge extraction.
Powered by Sarvam Vision, our new 3B vision-language model that delivers best-in-class accuracy for English and 22 Indian languages, outperforming Gemini 3 Pro and GPT-5.2 on global benchmarks.
sarvam.ai/blogs/sarvam-a…
Powered by Sarvam Vision, our new 3B vision-language model that delivers best-in-class accuracy for English and 22 Indian languages, outperforming Gemini 3 Pro and GPT-5.2 on global benchmarks.
sarvam.ai/blogs/sarvam-a…
6
42
262
9,734
Anna Upreti retweeted
After living in Delhi for my entire life, today was the first time the Delhi Airport felt like home. So happy to see @SarvamAI logo everywhere!
The energy feels different this week.
Every major power has had its defining AI moment. Now India begins writing its own chapter and @SarvamAI stands right at that inflection point.
Stay tuned for the mega drop.
IGI T3, New Delhi 📌
3
26
638
18,038
Anna Upreti retweeted
For 2 years, we asked ourselves the same question: How do you build a brand that is contemporary yet relatable for a billion Indians, a brand where culture and tech coexist? Our answer was to create a new reference point: Gateway. I’m incredibly proud of this team. More to do.
34
40
648
24,976
icon @AmandaAskell
Anthropic has entrusted Amanda Askell to endow its AI chatbot, Claude, with a sense of right and wrong on.wsj.com/4aC2kaW
0
0
6
855
Not your typical flashy AI demo but this is the real one: government partnerships, infra, and production deployment at scale for India
Drop 6/14: @SarvamAI is proud to announce a landmark in India’s sovereign AI journey through strategic partnerships with the Governments of Odisha and Tamil Nadu. The aim of these partnerships is to drive transformation by building at-scale compute, sovereign models, and the to drive AI adoption. Instead of going bottom-up with PoCs which find it hard to take off, these partnerships will go top down, with a state-wide mandate to deploy AI as a public utility, embedding intelligence across all departments.
4
1
39
3,029
Anna Upreti retweeted
Sovereign model update 12:
PT [████████████]
SFT [████████████]
RL [████████░░░░░░]
PT [████████████]
SFT [████████████]
RL [████████░░░░░░]
9
12
191
8,686
Anna Upreti retweeted
I was wrong about Sarvam.
When I wrote about them a year ago, I felt like the direction to train small "indic" language models was wrong. But boy, have they turned it around. They have the best text-to-speech, speech-to text, and OCR models for Indic languages, and that's e. The pricing is very reasonable. And the website is not only beautifully designed but dirt easy to use. They're filling a well needed gap in the ecosystem and doing things big labs will probably never focus on to the fullest extent (at least in the short term). I don't know anything about the business, but there's a lot to appreciate about what they've build technologically and I can't remember the last time I felt that way about software products coming out of India. Well done.
When I wrote about them a year ago, I felt like the direction to train small "indic" language models was wrong. But boy, have they turned it around. They have the best text-to-speech, speech-to text, and OCR models for Indic languages, and that's e. The pricing is very reasonable. And the website is not only beautifully designed but dirt easy to use. They're filling a well needed gap in the ecosystem and doing things big labs will probably never focus on to the fullest extent (at least in the short term). I don't know anything about the business, but there's a lot to appreciate about what they've build technologically and I can't remember the last time I felt that way about software products coming out of India. Well done.
153
476
4,684
658,445
Anna Upreti retweeted
The model can digitize a variety of documents -- from historical handwritten (Lincoln papers, Gandhi letters, Einstein notes) to old printed math with complex symbols or linotype era fonts (highlighted by @MohapatraHemant) in 22 Indian languages & English. More to come in vision.
This drop from @SarvamAI is really just so good I keep going back to it! Some of these fonts are so wild, it's magical they can even be transcribed. What's next team @SarvamAI?
3
9
56
6,834