Quick Take
- Sarvam AI is building a trillion-plus parameter model from scratch in India, unveiled at Epoch 2026.
- The frontier model targets coding, cybersecurity, science and simulation, and should arrive within six months.
- Its upgraded 105B model now costs $0.80 per million tokens, about 5.5X cheaper than GPT-5.4 Mini.
In This Article
Sarvam AI is building a trillion-plus parameter frontier model from scratch in India, cofounder Pratyush Kumar announced at the company’s first developer conference, Epoch 2026, in Bengaluru on July 30, 2026. The model is expected to go live within six months.
The Bengaluru-based AI unicorn made the reveal alongside its biggest product push yet, spanning models, an India-hosted inference service, and enterprise software. Kumar said the new model is being trained to compete in coding, cybersecurity, simulation and scientific research, moving Sarvam AI beyond Indian-language systems toward a full frontier-scale stack (as reported by trade press).
StartupFeed Insight
The real signal here is not the parameter count, it is the sequencing. Sarvam AI is anchoring the frontier model to a live inference business and a proven 105B voice engine, which means revenue can fund training rather than the reverse. Enterprises in BFSI, call centres and government should watch closely: if the cost story holds in production, the switching case becomes hard to ignore. StartupFeed expects Sarvam to ship at least one checkpoint of the trillion-parameter model, likely a smaller distilled variant, on Hugging Face before the end of Q1 FY2027, well ahead of a full frontier release. By Harshvardhan Jain.
Sarvam AI Trillion-Parameter Model: The Numbers
The Sarvam AI trillion-parameter model is a frontier-scale foundation model being trained entirely in India, placing the startup among a small global group attempting builds at this size. All figures below are from the company’s Epoch 2026 announcements. USD-to-INR conversions use the live rate of Rs 95.5 per dollar on July 31, 2026.
| Metric | Detail | Notes |
|---|---|---|
| Model scale | Trillion-plus parameters | Built from scratch in India |
| Focus areas | Coding, cybersecurity, science, simulation | Beyond Indian-language use cases |
| Expected launch | Within six months | Per cofounder Pratyush Kumar |
| Compute invested (existing models) | Nearly $20 Mn (Rs 191 Cr) | For three models incl. 105B, per Kumar |
| 105B token price | $0.80 (Rs 76) per Mn blended tokens | About 5.5X cheaper than GPT-5.4 Mini |
| Registered developers | More than 1 Mn | On Sarvam’s developer platform |
The most striking figure is the compute discipline. Kumar said nearly $20 Mn (Rs 191 Cr) in compute built three working models, a modest sum next to the sums frontier labs abroad spend, which frames the trillion-parameter bet as an efficiency play as much as a scale one.
About Sarvam AI
Sarvam AI is a Bengaluru-based artificial intelligence company founded in 2023 by Vivek Raghavan and Pratyush Kumar. an Indian ai building foundation models tuned for Indian languages and enterprise use, spanning text, speech, vision and voice agents. The company processes over 2 million daily interactions and was selected under India’s IndiaAI Mission. It raised a $234 Mn (Rs 2,235 Cr) Series B at a $1.5 Bn valuation, backed by investors including Lightspeed, Peak XV and Khosla Ventures.
Why is Sarvam AI building this model now?
Sarvam AI is building the model now to reduce India’s reliance on foreign AI systems, a goal it calls token sovereignty. Cofounder Vivek Raghavan framed the India-hosted inference service as central to serving more of the country’s AI compute needs on domestic infrastructure.
“We are building them from scratch to be competitive in coding, cybersecurity, simulation, science and more,” Pratyush Kumar, cofounder, Sarvam AI, said at Epoch 2026.
The timing follows Sarvam’s Series B and its selection under the government-backed IndiaAI Mission, with compute from data-centre operator Yotta and technical support from Nvidia. To guide the frontier effort, Sarvam appointed Devendra Singh Chaplot, part of the founding teams at Mistral AI and Thinking Machines Lab, as an advisor.
How cheap is Sarvam AI compared to rivals?
Sarvam AI’s upgraded 105B model costs $0.80 (Rs 76) per million blended tokens, which the company says is about 5.5X cheaper than OpenAI’s GPT-5.4 Mini at $4.50 and more than 11X cheaper than Google’s Gemini 3.5 Flash at $9. The 105B uses a mixture-of-experts design, activating only part of its parameters at once to cut compute cost.
Voice is where Sarvam pressed its cost edge hardest. The company argued that for teams building voice products in India today, nothing matches its price and scale, noting rival voice offerings are either unavailable or not at comparable scale. Sarvam also unveiled Bulbul V4 for expressive text-to-speech and Saras V4 for speech-to-text.
How does Sarvam AI compare to global labs?
Sarvam AI sits below the very largest global systems on raw scale but competes sharply on price and India focus. Frontier systems from OpenAI, Google and Anthropic are widely estimated to run into hundreds of billions, and possibly trillions, of parameters.
| Model | Price per Mn blended tokens | Positioning |
|---|---|---|
| Sarvam 105B | $0.80 (Rs 76) | India-hosted, voice-first |
| OpenAI GPT-5.4 Mini | $4.50 (Rs 430) | Global general-purpose |
| Google Gemini 3.5 Flash | $9.00 (Rs 860) | Global general-purpose |
What sets Sarvam apart is ownership of the full stack, from compute to inference to voice agents, tuned for Indian languages and priced for local scale.
What’s Next
Sarvam AI says the trillion-parameter model should arrive within six months, a demanding timeline given the hardware and talent needed. Watch for an intermediate checkpoint or a distilled variant to land first, plus wider rollout of its Samvaad voice-agent platform, which has handled over 325 Mn minutes of AI conversations. Can Sarvam turn stage ambition into reliable products at Indian scale?
Frequently Asked Questions
Have a tip? Write to us at editorial@startupfeed.in.
