Top News

AI unicorn Sarvam said it is building a trillion-parameter frontier AI model in India and launched an India-hosted inference service as part of its biggest product push yet to build an end-to-end AI stack spanning models, infrastructure and enterprise software. 

The Bengaluru-based startup made the announcements at its inaugural developer conference, Epoch 2026, alongside a range of launches across agentic AI, speech, vision and enterprise productivity. 

“We are very happy to announce that we are building a trillion-plus parameter model right here in India. We are building them from scratch to be competitive in coding, cybersecurity, simulation, science and more,” Sarvam cofounder Pratyush Kumar said at the event.

The startup said the new models are expected to go live in the next six months.

Sarvam also announced the opening of an office in San Francisco and appointed Devendra Singh Chaplot, who was part of the founding teams at Mistral AI and Thinking Machines Lab, as an advisor, signalling its plans to expand its global enterprise AI footprint. 

Among the biggest launches at the event was Sarvam Inference, an inference platform that serves frontier open-source AI models using infrastructure hosted within India.

The platform currently supports Sarvam’s own 105 Bn parameter model alongside open models such as GLM 5.2 and Gemma 4.

Sarvam said the offering is aimed at developers and enterprises that want access to frontier AI models while ensuring inference workloads remain within India, a capability that could appeal to enterprises and government organisations with data residency requirements. 

Cofounder Vivek Raghavan described the move as part of the startup’s broader push towards “token sovereignty”, or serving a larger share of the AI tokens consumed in India through domestic infrastructure rather than overseas cloud providers.

The startup also said it has developed an agentic optimisation system that automatically improves model performance and can deliver up to a 15X improvement in inference speed for certain models.

Sarvam added that its developer platform now has more than 1 Mn registered developers.

The startup also highlighted the pricing of its flagship 105 Bn parameter model, saying it costs $0.80 per 1 Mn blended tokens, compared with $4.50 for OpenAI’s GPT-5.4 Mini and $9 for Google’s Gemini 3.5 Flash. Based on these prices, Sarvam claimed its model is about 5.5X cheaper than GPT-5.4 Mini and more than 11X cheaper than Gemini 3.5 Flash. 

According to the startup, the model delivers competitive performance across instruction following, conversational AI and agentic workloads and is already powering voice AI deployments at scale. 

Beyond its infrastructure and frontier AI ambitions, Sarvam also unveiled a mix of upgraded AI models and new products across speech, vision, coding and enterprise workflows.

Sarvam Code: It is an AI coding agent designed for software development, machine learning, and cybersecurity workflows. The product uses separate planner, worker, and verifier agents to execute coding tasks. 

Sarvam Work: It is an enterprise AI agent that can analyse large datasets, access enterprise knowledge bases and carry out research and productivity tasks. It can be deployed through Slack, on-premise infrastructure or virtual private clouds. 

Sarvam said the products form part of Indus, its broader agentic AI platform that combines six offerings, including Samvaad, Work, Code, Content Agents, DocAgent and Inference, under a unified credit system. 

Speech And Vision Models Get Upgrades 

Saaras V4: It is the latest version of Sarvam’s speech-recognition model, with support for all 22 constitutionally recognised Indian languages and improvements in handling noisy environments, code-mixed conversations, Romanised text and multi-speaker interactions. 

Saaras V4 Multi-Speaker: Separate from Saaras V4, Sarvam has introduced a multi-speaker version built for conversations involving several people. It can identify and transcribe different speakers even when they interrupt each other or speak at the same time. It is also designed to work in noisy environments.

Bulbul V4: It is the latest version of Sarvam’s text-to-speech model. It builds on Bulbul V3, which was launched in February and focused on making AI-generated speech sound more natural. The new version gives developers greater control over expression. It can add emotions and laughter, change emphasis and even shift speaking styles midway through a sentence. Sarvam is positioning it for use cases such as voice agents, edtech, content creation, commentary and custom brand voices.

Sarvam Vision 2.0: The startup also upgraded its vision-language model with Vision 2.0, which is focused largely on reading and understanding documents. The first version was launched in February and could handle OCR, images, charts and complex tables.

Vision 2.0 improves handwriting recognition and table parsing. It also adds native information extraction from forms and support for all 22 Indian languages. Sarvam has also introduced Vision Edge, which can run locally on a computer. The technology is being used in projects such as digitising Odisha government land records.

Kivi: It is a desktop voice tool that lets users type, code and control their computers through speech. Built on Saaras V4, it supports Indian languages natively. The product is similar to voice-dictation tools such as Wispr Flow, but Sarvam is pitching Kivi more broadly around multilingual use cases and computer control.

Sarvam also announced a wider rollout of its Samvaad voice-agent platform, which was previously available largely to select enterprises and high-volume customers. The startup said Samvaad has handled more than 325 Mn minutes of AI-powered customer conversations over the past year.

Inc42 exclusively reported in June that Sarvam was preparing to expand the platform through a self-serve onboarding model for startups, SMBs, developers and individual users. Tata Capital is among the companies using Samvaad for multilingual customer interactions across its consumer lending business.

Voice AI remains Sarvam’s largest commercial business. Earlier, sources told Inc42 that conversational AI contributes nearly 80% of the startup’s estimated $12 Mn annual recurring revenue (ARR).

The post Sarvam To Build Trillion-Parameter AI Model, Launches India-Hosted Inference Service appeared first on Inc42 Media.