Sarvam AI
Indian foundation-model lab building open-weight LLMs and speech systems for Indian languages.
Sarvam AI builds open-weight language models trained in India, plus speech, vision, and enterprise products aimed at Indian languages. Sarvam-M (May 2025) is a 24B Apache-2.0 model post-trained from Mistral Small. Sarvam 30B and Sarvam 105B (open-sourced 6 Mar 2026) are reasoning MoE models trained from scratch on IndiaAI compute: 30B powers the Samvaad conversational platform, and 105B powers the Indus assistant. Weights are on Hugging Face and the IndiaAI AI Kosh; the API is OpenAI-compatible.
Why Sarvam AI matters
Sarvam is the clearest sovereign-AI option when the workload is Indian languages, Indic tokenization, or an open-weight model trained and served from India. If you are comparing Llama, Mistral, or DeepSeek for multilingual India deployments, Sarvam 30B and 105B belong on that short list.
Last reviewed: 25 September 2026
Best fit for
When architects typically choose Sarvam AI.
- Indian-language assistants and voice agents
- Open-weight reasoning models
- Sovereign or India-hosted AI stacks
- Efficient MoE deployment (30B)
- Agentic chat on Indus (105B)
Strengths
Qualitative snapshot for architects—not a public ranking.
- Indian languages★★★★★
- Open weights★★★★★
- Reasoning (105B)★★★★☆
- Deployment efficiency★★★★☆
- Global ecosystem★★☆☆☆
Quick facts
- Founded
- 2023
- Headquarters
- Bengaluru, Karnataka, India
- Ownership
- Private
- Open source
- Yes
- Enterprise
- Yes
- Flagship model
- Sarvam 105B
On DataAIHub
- 3Models
- 3Products
- 2Tools
- 3GitHub
- 2Research
- 9Guides
- 5Benchmarks
Company profile
- Founded
- 2023
- Headquarters
- Bengaluru, Karnataka, India
- Founders
- Vivek Raghavan, Pratyush Kumar
- Funding
- Private (~$41M seed and Series A, Lightspeed, Peak XV, Khosla)
- Ownership
- Private
- Country
- India
- Primary focus
- Open-weight foundation models, Indian-language LLMs, Speech and document AI, Sovereign AI for India
- Target users
- Developers, Enterprises, Public-sector teams, Researchers
- Revenue model
- API usage, Enterprise deployments
- Deployment
- Hosted OpenAI-compatible API plus Apache-2.0 weights for vLLM, SGLang, and Transformers
- Licensing
- Apache-2.0 open weights with separate API terms
- Open source
- Yes
- Cloud provider
- No
- Website
- https://www.sarvam.ai
- Confidence
- High
- Source coverage
- 8
Ecosystem
Competes with
- Mistral AI
Sarvam-M is a Mistral Small post-train; 30B and 105B compete as efficient open-weight alternatives.
- DeepSeek
Both ship open reasoning MoEs aimed at cost-sensitive self-host and API use.
- Meta
Llama is the default open-weight baseline Sarvam is compared with for Indic deployments.
Works with
- vLLM
Documented production serving path for Sarvam open weights.
- Sarvam 105B
Current flagship open reasoning model, used by Indus.
Recommended for
- Sarvam Models
How Sarvam-M, 30B, and 105B fit Indian-language and open-weight stacks.
- Large Language Models
Where an India-trained open LLM sits in the broader model landscape.
Often paired with
- Hugging Face
Sarvam-M, 30B, and 105B weights are published on the Hub.
How Sarvam AI evolved
Key moments in chronological order.
- Model
Document-intelligence model with structured extraction and Indic handwriting recognition. Separate from the text LLMs.
- Product
First Sarvam Epoch in Bengaluru covers new models, Sarvam Inference, Indus agents, and Sarvam Code.
- Open source
Sarvam 30B and 105B open-sourced
From-scratch reasoning MoEs trained in India. 30B (2.4B active) powers Samvaad; 105B (10.3B active, 128K context) powers Indus. Apache-2.0 weights on Hugging Face.
- Model
24B Apache-2.0 multilingual model, post-trained from Mistral Small, with an OpenAI-compatible API.
- Funding
About $41 million raised, with Lightspeed, Peak XV Partners, and Khosla Ventures among the backers named by Sarvam.
- Founded
Sarvam AI founded
Vivek Raghavan and Pratyush Kumar start a Bengaluru lab focused on AI for India.
Products
Foundation models
Related tools
Related research
- Open-sourcing Sarvam 30B and 105B
Original Paper
- Sarvam-M model card
Original Paper
GitHub
Related benchmarks
- MMLU
Broad knowledge and exam-style Q&A
- GPQA
Graduate-level science QA
- HumanEval
Python coding interview-style problems
- SWE-Bench
Real-world software engineering tasks
- Arena Hard
Hard chat preference eval
Related rankings
Related guides
Explore more companies
- Mistral AIFranceMistral AI builds Mistral Large 3, Small/Medium commercial tiers, and open-weight models, plus La Plateforme APIs and Le Chat. Mixtral remains an Apache-2.0 self-host option; it is not listed on Mistral serverless API pricing as of 2026-08-16. Regional Endpoints (EU or US) are generally available. Series D (€3B, 2026-09-08) funds sovereign open-weight infrastructure; a 2026-09-10 Cloudera partnership targets hybrid/air-gapped inference and custom models on customer data. Firefox Smart Window (beta, 2026-09-16) uses Mistral models in France and North America with zero data retention.
- DeepSeekChinaDeepSeek releases competitive open-weight models—including DeepSeek V4.1-Flash (live Flash, native multimodal), V4-Pro-0813 (still on deepseek-v4-pro after 2026-09-14), DeepSeek V3, and the reasoning-focused DeepSeek R1—plus APIs, chat, and DeepSeek Harness (MIT agent runtime, developer preview). It forced the industry to revisit cost/performance assumptions for frontier-class open models, especially for reasoning and coding workloads.
- Hugging FaceUnited StatesHugging Face provides the largest open platform for discovering, sharing, evaluating, and deploying machine learning models, datasets, and demos. It has become the default collaboration hub for the open-source AI ecosystem—backed by Transformers, Diffusers, PEFT, and managed inference. On 2026-09-03 NVIDIA agreed to acquire Hugging Face for $12.93B; until close (expected H1 2027, pending regulators) Hugging Face remains independent, and NVIDIA says the Hub will stay open to all model builders, clouds, and accelerators.
- AlibabaChinaGlobal technology group whose Qwen team releases competitive multilingual and multimodal foundation models for cloud and open-weight use.
- MetaUnited StatesMeta advances open multimodal research and releases the Llama open-weight model family used widely in self-hosted AI stacks. Muse Spark 1.3 (with Muse Code) is Meta’s paid Meta Model API path for agentic coding and multimodal agents. Combined with PyTorch, Llama still lets organizations train, fine-tune, and serve models without depending on a proprietary API.
- AI21 LabsIsraelFoundation model company known for Jurassic and Jamba models, plus enterprise generative AI products for writing and document workflows.