Sarvam’s ‘Shipmas’ Moment is Finally Here

Sector6  Feb 6, 2026 The bi-weekly newsletter by AIM that brings the biggest shifts shaping IT, AI, and GCCs. Sarvam's 'Shipmas' Moment is Finally Here Following the Sarvam Vision model, the startup is developing Bulbul 3, its next voice-based model, which is likely to be released this week. By Mohit Pandey For the past year, Sarvam has talked about sovereignty. This week, the Bengaluru-based startup began shipping at a pace that makes it appear more like a frontier AI lab, such as OpenAI or Anthropic.  The latest launch is Sarvam Vision, a three-billion-parameter vision-language model designed to read, understand, and extract information from documents in English and 22 Indian languages. The company is likely to announce a voice model named Bulbul soon.  On the surface, it resembles an OCR engine. In practice, it aims to do more. The model captions images, parses tables, understands charts, and converts messy scans into structured data. Less scanner software, more document intelligence. Sarvam's strongest claim is accuracy on Indic scripts, an area where global systems have historically struggled.  On its in-house Sarvam Indic OCR Bench, which includes 20,267 samples across scheduled languages, the model reports word accuracy of 95.91% for Hindi, 92.61% for Bengali, 93.42% for Tamil, 93.13% for Marathi, and 91.60% for Malayalam.  Lower-resource scripts such as Santhali and Dogri cross 80% accuracy in several cases. Global systems have historically struggled here because most are trained in English first and retrofit the rest. Most global OCR and vision-language systems are trained primarily in English and later extended to other languages. This approach often breaks down in Indian documents, where font layouts and scripts vary widely.  On the English olmOCR benchmark, Sarvam says it is competitive across math-heavy pages, tiny fonts, headers, footers, multi-column layouts, and tables. Table recognition touches 88.3%.  That detail matters. Indian workflows are dominated by forms, ledgers, invoices, and PDFs. When tables fail, automation fails.  The release also expands Sarvam beyond voice and text into multimodal AI.  Last week, the company rolled out Sarvam Audio ASR for a multilingual, multi-speaker speech recognition system.  Most ASR models work well on clean, single-speaker audio, but struggle with code-mixed language, interruptions, and overlapping speech, common in Indian customer support calls.  Sarvam says that in internal tests on the IndicVoices benchmark, its systems reported lower word error rates than GPT-4o-Transcribe and Gemini 3 Flash across unnormalised, normalised, and code-mixed speech. AIM Network Deep Dive >> While global giants focus on clean data, Sarvam is tackling the chaos of the real-world Indian market. From handling 8-speaker diarisation to enabling direct Speech-to-Action automation, Sarvam Audio is positioning itself as the critical infrastructure for the next wave of Indian digitisation. We dive deep into this in this episode of Front Page. Developers, Alert! AMD AI Engage is Here >> Join AMD AI Engage, an AMD-led initiative for developers and AI enthusiasts. Strengthen your AI skills, participate in challenges, and unlock opportunities for advanced certification experiences with AMD and industry experts. Register now. Let's look at some of the top news of this week >> Cognizant has partnered with Palantir Technologies to accelerate AI-led modernisation across healthcare and enterprise operations. OpenAI has announced Frontier, a new enterprise platform to help organisations build, deploy, and manage AI agents that can perform real-world work across business functions. AMD posted its strongest quarterly performance in years. Fourth quarter revenue jumped 34.1% year-on-year to a record $10.27 billion.  YouTube expands auto-dubbing feature to 27 languages, enhancing accessibility for creators. What's up with Sarvam? Sarvam is not slowing down. It began teasing Bulbul 3, the company's upcoming voice-based model, on its social media platforms. Put together, the stack now resembles a familiar playbook used by Frontier Labs. Build base models, layer APIs, cover modalities, and offer developers one platform. There is a clear strategy. Sarvam admits the system is not perfect. For vision, Bengali mistranslations show up. Instruction following weakens in low-resource languages. Similar things are there in the ASR model. Still, the gap versus general-purpose models appears large enough to matter. This finally increases expectations for the upcoming India AI Impact Summit, where the company is expected to announce its foundational LLM. Last year, the company announced one of its LLMs, Sarvam-M, built on top of Mistral's models, which received heavy backlash for lower rate of adoption amongst Indian developers. It seems the company has finally adopted a "building in public" approach, showcasing Indian products for praise and scrutiny alike. Sarvam was the first company selected under the IndiaAI Mission to build sovereign foundation models.  If Vision and Bulbul 3 land as promised, the company will no longer look like a niche Indic player but rather like India's answer to a full-stack AI lab. The GCC Advantage for US Mid-Sized retailers: A New Playbook for Scaling Analytics & AI in Retail AIM Research, in association with Systems Plus as research partner, has undertaken a comprehensive study to examine how GCCs—particularly in India—can help US mid-sized retailers bridge these gaps. The study finds that modern GCCs have evolved into strategic extensions of the enterprise, supporting core engineering, analytics, AI experimentation, and 24×7 operations. Click here to access the full report. Best Firm for Data Scientists – Showcase your leadership in data science and build a thriving workplace. Apply Now Join AIM Leaders Council – Connect with top data leaders and shape the future of AI. Check Eligibility Our Events – Be part of an exclusive event uniting North America's top data leaders. More Details AIM PeMa Quadrant – Discover the best in AI and data services with AIM's PeMa Quadrant. Explore Now GCC Explorer Unlock Exclusive Insights into 1400+ Global Capability Centers (GCCs) in India with AIM Research's Comprehensive Database.  Access GCC Explorer Stay ahead with these insights, and join us in driving the future of data science and technology. Your active engagement and participation are what make our community thrive. For Brand Collaborations, Contact us at info@aim.media To unsubscribe from future emails, simply click the "unsubscribe" link AIM. Unsubscribe © Copyright, 2026, AIM • 1st Floor, Sakti Statesman, Marathahalli – Sarjapur Outer Ring Rd, Green Glen Layout, Bellandur, Bengaluru – 560103