Sarvam AI Plans Trillion-Parameter Foundation Model
At its Epoch 2026 developer conference in Bengaluru, Sarvam AI announced plans for a trillion-plus-parameter foundation model to be built in India. According to the Free Press Journal, cofounder Pratyush Kumar said the model is intended for coding, cybersecurity, simulations, and scientific research, with availability expected within six months. The company also introduced updated voice models and new pricing for its Sarvam 105B model.
Sarvam AI announced plans at its Epoch 2026 developer conference in Bengaluru to build a trillion-plus-parameter foundation model in India.
According to the Free Press Journal, Sarvam cofounder Pratyush Kumar told attendees that the model is being trained from scratch for coding, cybersecurity, simulation, and scientific-research workloads. The publication reported that the company expects the model to arrive within six months. Economic Times, citing The Times of India, similarly reported that Sarvam announced the model as part of an effort to compete with systems from OpenAI, Google, and Anthropic.
Sarvam also disclosed operating details for its existing model portfolio. Economic Times reported that its current 100-billion-parameter model is focused on voice AI, conversational systems, English, dictation, and simulation-related tasks. Kumar said the platform had processed more than 325 million minutes of voice calls and had scored within four to five percentage points of a significantly larger model on one benchmark, according to Economic Times.
The Free Press Journal reported that Sarvam priced its upgraded Sarvam 105B model at $0.80 per million blended tokens. The company said that price was about 5.5 times lower than OpenAI's GPT-5.4 Mini at $4.50 per million tokens, and more than 11 times lower than Google's Gemini 3.5 Flash at $9 per million tokens.
The conference also introduced Bulbul V4, a text-to-speech model, and Saras V4, a speech-to-text model with multi-speaker separation for overlapping conversations, according to the Free Press Journal. For practitioners, comparable frontier-model announcements make deployment economics, language coverage, evaluation transparency, and inference reliability more consequential than parameter counts alone. Independent benchmarks and technical disclosures would be needed to assess how Sarvam's proposed model compares with established frontier systems.
Key Points
- 1Sarvam announced a trillion-plus-parameter model, but the sources describe its intended uses without providing architecture, training-compute, data, or independent benchmark evidence.
- 2Sarvam reported that its platform had processed more than 325 million minutes of voice calls, while its existing model is focused on voice and conversational workloads.
- 3Comparable frontier-model efforts make inference cost, evaluation transparency, multilingual performance, and operational reliability important criteria beyond parameter-count comparisons.
Scoring Rationale
A proposed trillion-parameter model from an Indian AI startup is notable for practitioners tracking new regional foundation-model suppliers and inference pricing. The announcement remains a roadmap rather than a released model, with limited technical specifications and no independent evaluations disclosed in the retrieved reporting.
Sources
Public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
