Sarvam AI Launches 24B Parameter Open-Source LLM for Indian Languages and Reasoning Tasks
You're reading Entrepreneur India, an international franchise of Entrepreneur Media.
Bengaluru-based AI startup Sarvam AI has introduced its flagship large language model (LLM), Sarvam-M, a 24-billion-parameter open-weights hybrid model built on Mistral Small. Designed with a focus on Indian languages and advanced reasoning capabilities, Sarvam-M is intended to power applications such as conversational agents, machine translation, and educational tools.
According to Sarvam, the model has been fine-tuned through a three-step process involving Supervised Fine-Tuning (SFT), Reinforcement Learning with Verifiable Rewards (RLVR), and inference optimisation. The SFT process involved crafting diverse, high-quality prompts to train the model in both general dialogue and complex reasoning. RLVR further improved its instruction-following and mathematical capabilities using custom reward engineering and curated datasets. Inference was optimised using FP8 post-training quantisation and techniques like lookahead decoding.
Sarvam-M has demonstrated strong performance in multilingual and reasoning benchmarks. It achieved a significant 86 per cent gain on a romanised Indian language version of the GSM-8K math dataset and showed average performance boosts of 20 per cent on Indian language benchmarks, 21.6 per cent on math, and 17.6 per cent on programming tasks. It outperforms Llama-4 Scout and is on par with models like Llama-3.3 70B and Gemma 3 27B, though it slightly lags in English benchmarks such as MMLU.
The model is accessible via Sarvam's API, its own playground, and available for download on Hugging Face. This release follows Sarvam's selection by the Indian government to build a sovereign LLM under the IndiaAI Mission, marking the first step in strengthening the country's domestic AI capabilities.
Hashtags
Try Our AI Features
Explore what Daily8 AI can do for you:
Comments
No comments yet...
Related Articles
Forbes
a few seconds ago
- Forbes
Exploring The Human Side Of AI
Sometimes in the push toward evaluating these big new technologies that we're seeing spring up, our responses to artificial intelligence seem somewhat schizophrenic. On the one hand, we're thinking a lot about hardware, data centers, energy needs, and how to support models. We have a lot of technical analysis involved in figuring out how AI applications are going to work logistically. There's also the other side of it – where I've repeatedly called for AI companies to hire an army of philosophers to figure out the human side. How will we interact with AI? How will it shape our social and emotional worlds, our family lives, our job lives, our lives in our communities? I heard an important voice on that second point recently in a TED talk by Noelle Russell who is the head of the AI leadership Institute and has a lot to say about progress toward accomplishing 'human' AI goals. An Interesting Journey Russell started at IBM, but went on to work at Red Hat and VMWare before landing at AWS, where she got some unique experience in building humanity into AI engines. Notably, Russell worked on the early Alexa project, which eventually became something so many of us are familiar with smart technology in our homes. Specifically, though, she worked on giving Alexa traits like mindfulness and kindness. Working on 17 different research models, Russell brought a unique perspective, and was sometimes challenged by her teammates. She talks about how as a mother and someone in the 'sandwich generation' also caring for a parent, she had a unique position on the team, and a different approach. Some, she said, thought of her as a 'buzzkill.' In terms of Alexa, more than a few early innovators didn't think AI would need, as Russell says, 'mindfulness in the kitchen,' but she was convinced that it would be an important aspect of this project – and it was, partly because we have begun to use Alexa for a lot more than just recipes. 'I realized that my unique perspective was not only changing life for the people that I work with, for the people in my own home, but it was actually changing the product for everyone, and that unless so many more people started providing this unique perspective, the product wouldn't serve everyone,' she said. Charting a Course: Aphorisms for Today Russell founded the AI leadership institute in 2016, and the organization offers certifications and all kinds of tools aimed at this central idea, that we need to promote mindfulness and deliberate user interface design for powerful LLMs. In interviews and presentations, she talks about the 'baby tiger' concept – AI is cute now, but what about when it matures? Russell also uses her own lamplighter metaphor that is markedly different from the one I often quote from Sam Altman. Altman's adage was that the modern knowledge worker, the equivalent of the old gas lamplighter, will be left behind in new technological revolutions, as workers re-skill and rise, phoenix-like, from their own ashes. Russell's lamplighter analogy is different. She says the lamplighter may go away, but future generations will still see the light of the lamp in question. That on its own provides more continuity between generations interacting with AI, and a generally positive outlook for crafting UX that works - for people. The Future - and History Presenting on that new job - of managing AI well - Russell noted a kind of unfortunate catch-22 or irony in modern communication. 'The number one skill (for AI management) is clarity of thought,' she said. 'You need to think about what you want, articulate what you want with words that a machine can understand. And here's the sad news. Over the last 10 years, we've gotten further and further away from that clarity of knowing who we are, and how we want to say things. We use emojis, right? We use these short words and (reductive symbols) and now these students are counting on us to be clearly, definitively identifying what we want to do, and how we want to build it.' She did offer a brighter note, though, in conclusion: Russell closed the TED talk with a quote from Hamilton, exhorting new generations to be bold in tackling things like AI design problems. 'I know that greatness lies in you,' she quoted, 'But remember from here on in… history has its eyes on you.' I guess history has its eyes on all of us. We should be mindful about what we create, and how we wade into our relationships with AI.
Associated Press
28 minutes ago
- Associated Press
T-ROC Global's Retail360 Wins 2025 TWICE VIP Award for Innovation in Retail Technology
MIAMI, FL - July 31, 2025 ( NEWMEDIAWIRE ) - The Revenue Optimization Companies (T-ROC Global), the leading provider of people and technology solutions for the global retail market, announces its Retail360 product has won this year's TWICE VIP Awards 2025 in the Innovation in Retail Technology category. Developed in-house by T-ROC Global, Retail360 is a fully integrated, proprietary ecosystem engineered to empower field reps, delight shoppers, and drive measurable business results. The revolutionary AI-powered platform is redefining in-store performance. 'Retail360 represents the future of retail execution - where cutting-edge AI meets real human insight to deliver smarter, faster, and more personalized experiences at every level of the store,' said Brett Beveridge, CEO and Founder of T-ROC Global. 'This award is a testament to the incredible innovation and hard work of our team, and to our unwavering mission to transform retail through technology that truly works for people.' The TWICE VIP Awards celebrate and recognize the best features, design and value that new products are bringing to consumers. Voted upon exclusively by TWICE magazines readers and subscribers, the leading voice in the consumer electronics industry, this achievement highlights which companies are truly going above and beyond to deliver value to consumers. Retail360 features six core AI-driven solutions: Optinomics, WorkLink, NextStop, VisionIQ, SmartPix, and Navigator. Retail360 manages every step of the retail execution journey, from market planning to post-visit analysis. It uses generative AI, machine learning, image recognition, and real-time competitive data to deliver clarity, consistency, and control across thousands of retail locations. From optimizing territory coverage to providing visual proof of shelf compliance, the platform surfaces data transforming it into actionable insight. To learn more about Retail360 and how T-ROC Global is revolutionizing the retail landscape, visit T-ROC Global can also be found on Facebook, X, Instagram, and LinkedIn. About The Revenue Optimization Companies (T-ROC Global) T-ROC Global is a retail branding and consulting partner that supports companies in navigating through today's retail shopping experience, redefining the power of people and technology. T-ROC Global offers a unique combination of people-based services, applications, technology management, mystery shopping programs, actionable market research and competitive insights that support the complex needs of assisted selling. T-ROC Global's expertise and next-generation technology is delivered by a team that's all in to drive sales, optimize performance and deliver measurable ROI for businesses every single day. Media Contact: Tyler Sminkey, (786) 390-8510, [email protected] View the original release on
Associated Press
28 minutes ago
- Associated Press
SOUNDRAW Launches Global 'Beat the Future' Contest with $2,500 Prize Pool and Free AI Music Access
AI music platform SOUNDRAW invites global artists to enter its 'Beat the Future' contest by customizing AI-generated beats. Winners will share $2,500 in prizes, plus spotlight features and SOUNDRAW subscriptions. Tokyo, Japan, July 31, 2025 -- SOUNDRAW, an ethical AI-powered music generator built by a team of professional producers in Japan, has announced the launch of 'Beat the Future,' a global contest designed to spotlight emerging artists across the music industry. The competition is open to producers, singers, and rappers worldwide, encouraging participants to create original music using SOUNDRAW's platform. The contest aims to demonstrate that with the right tools, anyone—regardless of background or experience—can create high-quality music. SOUNDRAW's platform enables fast, genre-specific beat generation using features such as stem downloads, instrument swapping, track length customization, and genre mixing. All music is fully copyright-safe, as it is generated from in-house content rather than scraped data. 'Beat the Future' builds on the success of SOUNDRAW's recent RAW RAP CHALLENGE series, which featured 10 emerging rappers performing on the popular show On The Radar. Contest Details: Prizes: The contest is free to enter and open to creators of all skill levels. SOUNDRAW is widely used by independent musicians, content creators, and YouTubers for generating royalty-free music tailored to their needs. 'With Beat the Future, we want to empower bedroom producers, TikTok stars, and undiscovered talents to show off what they can do with a little creativity and a powerful tool,' said Anita Baumgärtner, Head of Marketing at SOUNDRAW. To learn more or submit your entry, visit the official contest page at About SOUNDRAW SOUNDRAW is an AI-powered music generation platform designed to support artists, creators, and musicians in producing original tracks quickly and ethically. Unlike many AI tools, SOUNDRAW generates music from proprietary, in-house content, ensuring full copyright safety for commercial use. Contact Info: Name: SOUNDRAW Email: Send Email Organization: SOUNDRAW Address: Yoyogi 5-63-4 3F, Shibuya ward, Tokyo, Japan 151-0053 Website: Release ID: 89166128 If you encounter any issues, discrepancies, or concerns regarding the content provided in this press release, or if there is a need for a press release takedown, we urge you to notify us without delay at [email protected] (it is important to note that this email is the authorized channel for such matters, sending multiple emails to multiple addresses does not necessarily help expedite your request). Our expert team will be available to promptly respond within 8 hours – ensuring swift resolution of identified issues or offering guidance on removal procedures. Delivering accurate and reliable information is fundamental to our mission.
