Pinned Tweet
๐ฎ๐ณ Introducing Anuvani-LID: A new state of the art in spoken language identification for Indian languages.
Before AI can understand what we say, it must first recognise the language we speak.
Today, the Central Institute of Indian Languages (CIIL), Government of India, unveils Anuvani-LID. It is our first open-source AI model and the first release in the Anuvani family of speech AI models.
India's linguistic diversity needs models that can tell closely related languages apart and work reliably outside the lab.
๐ Anuvani-LID ranked #1 among 16 leading AI systems in our SUITE-42 evaluation. Ahead of Gemini 3.5 Flash, Sarvam Saaras v3, OpenAI's GPT-audio, Whisper-large-v3, Meta's MMS-LID family and others.
๐ฅ 0.9125 mean Macro-F1 across 37 benchmark conditions (Gemini 3.5 Flash: 0.8523; Sarvam Saaras v3: 0.8372)
โก 24 percentage points ahead of Gemini 3.5 Flash on one-second speech
๐ 0.897 Macro-F1 across all 23 supported languages, vs 0.713 for the strongest competing system
๐๏ธ 22 constitutionally scheduled Indian languages + Indian English in one unified model
๐ง Trained on 54,643 hours of speech (31M+ utterances) with a compact 115.6M-parameter Conformer
It's built for real-world speech: short utterances, background noise, reverberation and closely related languages.
๐๐ป๐ฑ ๐๐ฒ'๐ฟ๐ฒ ๐บ๐ฎ๐ธ๐ถ๐ป๐ด ๐ถ๐ ๐ผ๐ฝ๐ฒ๐ป ๐๐ผ๐๐ฟ๐ฐ๐ฒ. Released under Apache 2.0, with support for PyTorch, ONNX, NVIDIA Triton and mobile inference:
โข ๐ก๐ฉ๐๐๐๐ ๐ง๐ฟ๐ถ๐๐ผ๐ป ๐๐ป๐ณ๐ฒ๐ฟ๐ฒ๐ป๐ฐ๐ฒ ๐ฆ๐ฒ๐ฟ๐๐ฒ๐ฟ: Production-ready serving kit for scalable, high-performance inference.
โข ๐ ๐ผ๐ฏ๐ถ๐น๐ฒ ๐๐ป๐ณ๐ฒ๐ฟ๐ฒ๐ป๐ฐ๐ฒ ๐๐ถ๐: Cross-platform toolkit for language identification on mobile and edge devices.
โข ๐๐ผ๐ผ๐ด๐น๐ฒ ๐๐ผ๐น๐ฎ๐ฏ: Beginner-friendly notebook to explore, test and experiment with Anuvani-LID.
โข ๐๐๐ด๐ด๐ถ๐ป๐ด ๐๐ฎ๐ฐ๐ฒ: Open model weights and resources for seamless integration.
๐๐ฟ๐ผ๐บ ๐ฟ๐ฒ๐๐ฒ๐ฎ๐ฟ๐ฐ๐ต ๐ฒ๐
๐ฝ๐ฒ๐ฟ๐ถ๐บ๐ฒ๐ป๐๐ ๐๐ผ ๐ฝ๐ฟ๐ผ๐ฑ๐๐ฐ๐๐ถ๐ผ๐ป-๐๐ฐ๐ฎ๐น๐ฒ ๐ฑ๐ฒ๐ฝ๐น๐ผ๐๐บ๐ฒ๐ป๐๐, ๐ณ๐ฟ๐ผ๐บ ๐ฐ๐น๐ผ๐๐ฑ ๐๐ฒ๐ฟ๐๐ฒ๐ฟ๐ ๐๐ผ ๐บ๐ผ๐ฏ๐ถ๐น๐ฒ ๐ฑ๐ฒ๐๐ถ๐ฐ๐ฒ๐. ๐๐ป๐๐๐ฎ๐ป๐ถ-๐๐๐ ๐ถ๐ ๐ฏ๐๐ถ๐น๐ ๐๐ผ ๐บ๐ฎ๐ธ๐ฒ ๐๐ป๐ฑ๐ถ๐ฎ๐ป ๐น๐ฎ๐ป๐ด๐๐ฎ๐ด๐ฒ ๐๐ ๐ฎ๐ฐ๐ฐ๐ฒ๐๐๐ถ๐ฏ๐น๐ฒ ๐ฒ๐๐ฒ๐ฟ๐๐๐ต๐ฒ๐ฟ๐ฒ.
Recognising the spoken language is the first step for voice assistants, ASR, translation and accessibility. Anuvani-LID is the first milestone in CIIL's Anuvani speech AI ecosystem, with more models and open-source releases on the way.
Built for India's languages. Open to the world. ๐ฎ๐ณ
๐ ๐๐
๐ฝ๐น๐ผ๐ฟ๐ฒ ๐๐ต๐ฒ ๐ผ๐ณ๐ณ๐ถ๐ฐ๐ถ๐ฎ๐น ๐น๐ฎ๐๐ป๐ฐ๐ต ๐ฎ๐ป๐ฑ ๐ฟ๐ฒ๐๐ฒ๐ฎ๐ฟ๐ฐ๐ต:
ai.ciil.org/release/anuvani-โฆ
๐ค ๐๐ผ๐๐ป๐น๐ผ๐ฎ๐ฑ ๐๐ต๐ฒ ๐บ๐ผ๐ฑ๐ฒ๐น ๐ผ๐ป ๐๐๐ด๐ด๐ถ๐ป๐ด ๐๐ฎ๐ฐ๐ฒ:
huggingface.co/CIIL/Anuvani-โฆ
๐ ๐ฅ๐ฒ๐ฎ๐ฑ ๐๐ต๐ฒ ๐ฟ๐ฒ๐๐ฒ๐ฎ๐ฟ๐ฐ๐ต ๐ฝ๐ฎ๐ฝ๐ฒ๐ฟ:
arxiv.org/abs/2610.20263
#AI #SpeechAI #IndianLanguages #OpenSource #LanguageTechnology #CIIL
At this pace Google Deepmind will speed run the whole periodic table before Christmas
RSI achieved ๐ซก
2
29
Stephen Fernandes retweeted
before you disaggregate prefill and decode, do one calculation.
cache transfer time = kv cache size / link bandwidth.
disaggregation puts prefill and decode on separate gpu pools so they scale and schedule independently, and neither stalls the other. in exchange, every request ships its entire kv cache from one pool to the other.
if that transfer time is anywhere near your prefill time, the split is eating its own win. on fast in node links it can be tiny. across nodes on a slower fabric, it can be the whole latency story. measure the link first, benchmark second.
People who never have launched a slurm job,
people who dont even know what .einsum does in pytorch
people who have never written a WGMA kernel
are throwing shade at a monumental startup which India essentially needs in the coming years
A startup that is handling the most critical infra of many top institutions and MNCs in the country.
In Psychology this is called : " Schadenfreude " feeling intense joy and satisfaction in hoping to see others fail.
enjoy the clout you clown, post this you would just be crawling back to your meaning life whoch serves no existence
How is this not a fraud?
India's leading AI lab worth $1.4 billion got funded and GPU to build a sovereign model in India by the Government.
They have now stopped innovating.
..and using the same GPU to sell Chinese model openly.
This is absolutely rubbish and mockery.
6
19
1,044
Stephen Fernandes retweeted
People who haven't read a single research paper.
Are now judging how much money it takes to build frontier models.
They throw around terms like distillation and trillion parameter.
Without knowing what those terms mean.
This is why innovation is hard. Not only do you have to manage the models.
You have to manage the public perception, including the uninformed public.
For all the people taking a jab at @sarvam. dig the deepest hole you possibly can when the 1T model is released you are going to need it.
we are in the 2nd loop of the sarvam M cycle ๐คซ
Introducing Jev @typesafeai powered jobs-api
Try it at treg.to/jobs-api
- One API to any jobs data
- $0.03 to run a full job hunt
- Real salary & Hiring manager data
- Plugin to any agent for your dream job
Fully open source, 0% markup
Git Repo below ๐
for almost a couple of decades, this large software engineering industry put food on the table for millions of people.
sadly with the advent of claude code and codex this large monolithic machinery has hyper optimized itself to remove the most unoptimized cogs in it
490
Claude opus and fable are getting insanely great at software engineering to the point where its actually scary.
literally anybody can write a major component in IT that would have taken millions in engineering costs.
> it took us 4 days and 500$ in tokens
> we beat SGLang by 15%
> it took 6 prompts
> 3 out of those 6 were:
(a) "carry on"
(b) "proceed"
(c) "push to github"
and... we did it from SCRATCH (git init)
context: it has recently become a very clear industry trend to build a custom inference engine for each open source model that comes out... a paradigm made popular by a WAVE of papers that came out earlier this year, like hazy research's megakernels and MetaInfer
the usual argument against traditional inference engines (like vllm / sglang) is that they, like all conventional swe products, use abstractions. the use of abstractions is synonymous with performance overhead (triton vs cuda, python vs c++, etc). combine this with increasing hardware complexity, increasing parallelization strategies, increasing kernel specialization techniques... and you can make a very clear (but very FLAWED) argument for building an engine customized purposely for one model.
why flawed? it is true that traditional inference engines waste cpu time running generic abstractions. sure. but the gpu kernels (the functions run on the gpu) are almost always longer, and so the wasted cpu time is not the bottleneck. if you know your shapes, and run cuda graphs ahead of time (and you almost always do), this is NEVER a problem.
so why would you build your own engine? why are people RAVING about this?
so we tested it out. we built one for DeepSeek V4.1 Flash. The goal was two-fold:
1) is it feasible to do so (spoiler: yes)
2) is it useful to do so (spoiler: also yes...but for different reasons)
we built it in rust (in retrospect, this did not really matter), and it took us 6 prompts and cost 500$ end to end.
unshackled from working within a pre-existing codebase or inheriting code debt, we:
- implemented custom kernels (GEMV, MoE, and AllReduces) specifically for DSV4.1 at our batch size
- we were able to put an entire spec dec step in one graph, and verify across 7 tokens at that
- we were able to mix deepseek's 4 residual streams in 1 fused kernel instead of 80
- fused the routing so that picking and running the expert happened in a single step (saving one write and one read i/o from each expert)
iterating on the model was cheap to do in a custom sandbox-like engine, and we were able to go much faster than a traditional dev cycle. we were then able to port every single thing into sglang, and matched performance
imo, it is this easy iteration that makes this paradigm useful. little else matters.
and, ofc, we've open-sourced the engine and upstream(ed/ing) the patches to @sgl_project
2
149
Lets Go !!!!!
37
Those who have viewed Lex's vlog featuring Jensen in Taiwan would undoubtedly recognize the profound humility, gratitude, and generosity Jensen consistently shows toward others. It is deeply painful and disheartening for me to see him depart.
NVIDIA letting Jensen go is a huge mistake, for America and for humanity. It's Apple firing Steve Jobs all over again, but much bigger, given the role of AI in defining the future of our world.
I've been keeping quiet on this out of respect for Jensen's family and team, who I've met and who are amazing humans, and because I personally hate infidelity. So my heart goes out to them in this difficult personal family moment.
But I can also say that Jensen is human, and makes mistakes like all of us do. Not that I matter at all, but I've known Jensen for a long time and have gotten the chance to spend a lot of time with him, behind the scenes, away from mics and cameras. I have seen him show kindness, humility, care to everyone around him, including strangers, including from the poorest most struggling parts of society, treating all the same with the same thoughtfulness and attention, because they are fellow humans and not because he has something to gain from them. That shows rare character, to maintain humility and humanity through all of it.
Another data point is what people who know him well have said privately to me about Jensen over the years. I've never had anyone tell me anything but how awesome he is or how much they love him. Probably over a thousand people at this point. This includes hundreds of engineers, CEOs, colleagues, friends, etc.
Also, since there was some spy talk. Jensen is as proud an American as it gets. I'm in shock frankly at all of this, including the mass hysteria around it. I love America, and I think NVIDIA letting Jensen go is terrible for America.
I stand with Jensen.
Love you all โค๏ธ
1
111




















