Arjun Reddy

@SarvamAI @SarvamForDevs @harsh_m121 @VinayakGavariya would love your thoughts

Hi @SarvamAI @SarvamForDevs We used TurboQuant to shrink Sarvam30B ~3.5 times smaller while losing only 2.5% performance compared to BF16

Compress the model. Keep the intelligence. Sarvam built a remarkable 32-billion parameter model for a billion Indian language speakers. Our job was making it run anywhere. 64 gigabytes is a lot of weight to carry. We applied TurboQuant. The core insight is simple. Rotate the weights so every coordinate carries equal information. Now each number looks like every other number. No outliers. No wasted bits. Quantize each to one of eight optimal values. Three bits instead of sixteen. Add two tiny scale factors per block to recover the fine detail. The experts — 90% of the model — got the aggressive treatment. Attention weights got slightly more room. Router weights stayed untouched. Routing is a discrete decision. You don't approximate discrete decisions. 64 gigabytes became 18.6. One GPU instead of four. Cosine similarity above 0.99. The weights changed. The intelligence didn't. Specific knowledge plus leverage. Know which weights can compress. Apply the math. Ship it

@SarvamForDevs @VinayakGavariya @SarvamAI You’re the best! Now gonna roll out Sarvam Voice Models in our Agent for all Indian languages, OpenClaw for India! Thank you @SarvamAI @SarvamForDevs

@junafinity @VinayakGavariya @SarvamAI Hey Arjun! 10k credits have been added to your account. Hope these credits are used to make the experience better. The agent looks solid excited to see more people discover it. All the best as you scale! 🛠️✨

This sounds like an April Fool’s, but it’s not. If you’re building with @SarvamAI, reply with: • The email you’ve used to log in to dashboard.sarvam.ai • A short note on your use case We’ll add ₹10,000 in credits before April 1st ends

Google compromised BTC & Blockchain Most crypto lives on Elliptic Curve Cryptography. It has been a vault for a decade. Quantum computing is the crowbar. Shors algorithm turns a public key into a private one. If the math is exposed, the money is gone. Understand the distinction: Your public address is a hash—a secure mask. As long as the assets sit still, the mask holds. But to move them, you have to reveal the raw public key to the mempool to prove ownership. This is the window of vulnerability. For a few minutes, you are standing in a dark forest with a spotlight on your head. A quantum attacker does not need your address; they just need those raw coordinates to solve the math and front-run your transaction. Google quantified this quantum gap. They are pushing for Post-Quantum Cryptography. If you want to build a Quantum Safe Crypto Project, here’s how you do it: Algorithms like ML-DSA are not just upgrades; they are the new armor. Practical fixes like commit-delay-reveal keep the raw key hidden until the network is ready to lock the door. In digital markets, there are no bailouts. You either master the tail risks or they master you. The new rules: 1. The address is a mask; the key is the face. Never show your face in an unsecure room. 2. The mempool is the danger zone. Minimize the time between broadcast and confirmation. 3. PQC is the only exit. If the math can be broken, the system is already dead. Wealth is permissionless, but it is not easy. It requires the hardest math to survive the future. Build accordingly.

@Apple has stopped selling the 512GB RAM variant on their M3 Ultra Mac Studios. RAM shortage or new version coming soon?

@SarvamForDevs Someone was using LiteLLM?

⚠️ Earlier today, Sarvam APIs experienced a period of downtime affecting all endpoints- STT, TTS, Translation, LLM, and Vision. Our team identified the issue and services are now recovering. Most APIs are back online, though some users may still see elevated latency or intermittent errors as systems stabilize. The API Dashboard (dashboard.sarvam.ai) is currently still down as part of the recovery process. API keys continue to work as expected. We're actively monitoring and will keep you posted. For real-time updates → status.sarvam.ai Apologies for the disruption, and appreciate your patience while we get everything fully back to normal.

@intel @IntelAI we are 50% done implementing @GoogleAI ‘s Turbo Quant for OpenVINO

@nabilajamal_ Did that magistrate face the same consequences as the policemen?

#TamilNadu In June 2020, Jayaraj and his son Benicks were arrested for allegedly keeping their mobile shop open past lockdown hours. That charge was later proven false They were tortured overnight at a police station in Thoothukudi. A magistrate remanded them to jail, ignoring their visible injuries. Both died within days. Now, nearly six years later, all nine police officers accused of their deaths have been convicted Sentencing is scheduled for March 30. Their family demands death penalty!

@TheAhmadOsman Will it have vision?

Here we go Confirmation from MiniMax’s Head of Engineering on what I have been saying: MiniMax-M2.7 weights will be opensourced within the next couple of weeks

@OfficialINDIAai @ai4bharat @SarvamAI @SarvamForDevs @caleb_friesen @Analyticsindiam

@SarvamForDevs @SarvamAI would love it if you can match this helping out startups with Sarvam credits

@libsoftiktok I thought it died at the end of “Weapons” movie

OH MY GOD Two foster kids in Australia were placed in the same house as a TRANSGENDER SERIAL KILLER People should go to prison for this

@AmazingSights @tacobell for insects?

Chemical defense is a game-changer for the bombardier beetle. Watch how it sprays acid at a striped spider. [📹 jackmindfresh1]

@SarvamForDevs Congratulations team! Is your 30B model compatible for speculative decoding of your bigger model?

everyone behind Sarvam Model APIs.

@pratykumar Thank you for your gift to the OSS community @pratykumar is your 30B model compatible for Speculative Decoding of your 105B model?

📢 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…