Short answer: Sarvam AI is a Bengaluru-based company building an Indian-language AI stack: foundation models, speech recognition, translation, text-to-speech, document processing, developer APIs and the consumer assistant Indus. It is not simply an Indian copy of ChatGPT. Its central proposition is “sovereign AI”—greater Indian control over models, data, infrastructure and public-service interfaces—combined with practical support for India’s many languages and code-mixed speech.
What is Sarvam AI?
Sarvam AI was founded in August 2023 by Dr. Vivek Raghavan and Dr. Pratyush Kumar. Raghavan is associated with India’s digital-public-infrastructure efforts, while Kumar has worked on Indian-language AI research, including work connected with AI4Bharat. The company describes its mission as building AI for India and developing sovereign capabilities across models, infrastructure and applications.
That makes Sarvam several things at once:
- An AI company: founded and headquartered in Bengaluru.
- A model laboratory: developing language, speech, translation, text-to-speech and document models.
- A developer platform: offering APIs and SDK access for chat, audio, translation and document workflows.
- A consumer-product maker: its Indus assistant provides a ChatGPT-like interface to Sarvam models.
Keep those labels separate. Sarvam AI is the company; Sarvam-105B is a model; Sarvam API is the developer platform; Indus is the consumer-facing product.
Why Indian-language AI is difficult
India’s language environment is not solved by adding a translation button to an English-first chatbot. The country has 22 constitutionally recognized scheduled languages, multiple scripts, substantial dialect and accent variation, informal spelling, and widespread code-mixing such as Hinglish or Tamil-English. Real deployments also involve low-bandwidth connections, telephone-quality 8 kHz audio, background noise and speakers switching languages mid-sentence.
#1 Best Overall
Sarvam’s documentation specifically emphasizes code-mixed input, transliteration, native numerals and telephony audio. Quality still has to be measured language by language: availability of a language does not prove equal speech recognition, translation, text-generation or text-to-speech performance in that language.
What “sovereign AI” means
For Sarvam, sovereignty is about more than owning a chatbot. It can include where models are trained and run, who controls data and deployment, whether public services depend on foreign providers, and whether systems can be adapted to Indian languages, laws and administrative workflows.
The Government of India selected Sarvam in April 2025 to build an indigenous large language model under the IndiaAI Mission, with dedicated compute resources. A February 2026 government overview also listed Sarvam among organizations selected for the foundation-model initiative. That support does not make every Sarvam product a government system, guarantee privacy, or make Sarvam the only Indian foundation-model company. Government materials describe several participants, including BharatGen, Gnani and Soket AI.
Sarvam’s IndiaAI announcement · Government overview · IndiaAI foundation-model organizations
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Rank #2
Sarvam’s products and models
| Product or model | Main job | What it means for users |
|---|---|---|
| Sarvam-105B | Chat, reasoning, coding and agentic workflows | Flagship general-purpose language model |
| Sarvam-30B | Smaller language model | Potentially easier to deploy at lower cost; current checkpoint and hardware requirements should be verified |
| Saaras v3 | Speech-to-text, translation and transliteration | Processes Indian and code-mixed speech; results depend on accent, noise and audio quality |
| Bulbul v3 | Text-to-speech | Generates spoken output in Indian languages |
| Mayura | Focused translation | Translates selected Indian-language pairs |
| Sarvam-Translate | Broader translation | Wider coverage, with quality varying by language direction |
| Sarvam Vision | OCR and document intelligence | Extracts text, tables and structured data from documents |
| Indus | Consumer chat | ChatGPT-like interface powered by Sarvam’s models |
| Sarvam API | Developer access | REST and SDK access to language, audio, translation and document services |
Sarvam-105B explained
Sarvam describes Sarvam-105B as a 105-billion-plus-parameter mixture-of-experts model trained from scratch on 12 trillion tokens. Its published design includes 128 sparse experts and Multi-head Latent Attention. The documentation says it is available through an OpenAI-compatible chat-completions API and identifies the model under an Apache 2.0 license.
What mixture-of-experts means
A mixture-of-experts model does not necessarily activate all 105 billion parameters for every token. A routing system selects a subset of experts for each input. That can reduce computation compared with a dense model of the same total size, but actual latency and cost still depend on hardware, serving software, context length and batching. “105B” should not be read as “105B active parameters” unless Sarvam publishes that figure for the specific checkpoint.
How to read Sarvam’s benchmark claims
Sarvam’s documentation reports scores including 98.6 on Math500, 88.3 on AIME 25, 96.7 on AIME 25 with tools, 49.5 on BrowseComp and a 90% pairwise win rate for Indian-language performance. These are company-reported figures, not independent proof that Sarvam is better than ChatGPT, Gemini or every open model.
A serious evaluation should identify the exact checkpoint, prompts, tool settings, comparison models and evaluator. It should also test production conditions such as code-mixed speech, noisy calls, names, addresses and regional scripts. Benchmark capability does not by itself establish factuality, safety or reliability.
Sarvam-105B specifications and reported benchmarks
What is Indus?
Indus is Sarvam’s consumer-facing assistant, while Sarvam-105B is the underlying model and Sarvam API is the developer route. Sarvam introduced Indus in limited beta on February 20, 2026, warning that access could be constrained by a waitlist or available compute.
That beta history matters. Indus is useful for trying Sarvam’s Indian-language experience, but readers should check current access, features, uptime, mobile availability and pricing before treating it as a fully mature replacement for ChatGPT or Gemini.
Sarvam versus ChatGPT and Gemini
This is not a symmetrical comparison. ChatGPT and Gemini are mature global consumer ecosystems that typically combine chat, multimodal input, browsing or search, coding tools, applications, integrations and extensive safety operations. Sarvam combines Indian-language models with speech, translation, document APIs, sovereign-deployment ambitions and an emerging consumer interface.
| Criterion | Sarvam | ChatGPT | Gemini |
|---|---|---|---|
| Indian-language specialization | Core product focus, including speech and code-mixing | Check language and feature support for the exact use case | Check language and feature support for the exact use case |
| Consumer assistant | Indus, with limited-beta history | Mature consumer product | Mature consumer product |
| Open-weight availability | Sarvam documentation identifies Sarvam-105B as Apache 2.0; verify checkpoint terms | Product/model specific | Product/model specific |
| Speech and TTS | Saaras and Bulbul | Product/API specific | Product/API specific |
| Sovereign deployment | Central positioning | Contract and region dependent | Contract and region dependent |
| Best initial audience | Indian-language developers, enterprises and public services | Broad global users and software teams | Broad users, especially those in Google’s ecosystem |
When Sarvam may fit better
- Indian-language speech recognition, translation or text-to-speech is central.
- The application must handle code-mixed or telephone-quality audio.
- A public-sector or enterprise buyer needs deployment control and local integration.
- A developer wants an OpenAI-compatible interface or an open-weight option.
- Rupee-denominated usage pricing and India-focused support are valuable.
When a global platform may fit better
- The work is primarily broad English-language assistance.
- You need a mature multimodal consumer product, browsing, connectors or coding-agent ecosystem.
- Your organization requires established global administration and independently published reliability comparisons.
There is no evidence here for declaring one universal winner. The right comparison is task, language, deployment and compliance specific.
Free tools Windows power users keep installed
One-click scans. No signup required.
What developers can build
Voice public services
A service could combine Saaras for transcription, Sarvam-105B for dialogue, Mayura or Sarvam-Translate for language conversion, Bulbul for spoken replies and retrieval over official documents. High-risk answers still need validation and human escalation.
Call centers
Relevant capabilities include transcription, diarization, translation, summarization and agent assistance. Test overlapping speakers, background noise, accents, names, addresses, account numbers and mixed-language utterances before deployment.
Document workflows
Sarvam Vision or document APIs may help digitize forms, invoices and multilingual scans. Validate output on skewed pages, stamps, handwriting, tables, multi-column layouts and similar-looking characters before writing extracted values into a database.
Education and media
Potential uses include local-language tutoring, textbook translation, accessibility, subtitles, dubbing, podcast transcription and regional publishing. Human review remains important because generated explanations can be wrong, curricula can be mismatched and voice rights or consent may restrict commercial use.
Best Value
Enterprise agents
An agent is more than a chatbot. Production systems need narrowly scoped tool permissions, retrieval grounding, audit logs, deterministic checks, prompt-injection defenses, human approval for sensitive actions and recovery paths when tools fail.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.API pricing and access
The following prices were listed on Sarvam’s pricing page on August 16, 2026. Prices, taxes, account limits, regional terms and enterprise discounts can change; billing units differ by service.
| Service | Listed price | Billing unit or note |
|---|---|---|
| Sarvam-105B input | ₹4 | Per million input tokens |
| Cached input | ₹2.50 | Per million cached input tokens |
| Sarvam-105B output | ₹16 | Per million output tokens |
| Speech-to-text | ₹30 | Per hour |
| Speech-to-text with diarization | ₹45 | Per hour |
| Speech-to-text and translate | ₹30 | Per hour |
| Speech-to-text, translate and diarization | ₹45 | Per hour |
| Sarvam Translate V1 | ₹20 | Per 10,000 characters |
| Mayura V1 | ₹20 | Per 10,000 characters |
| Transliteration | ₹20 | Per 10,000 characters |
| Language identification | ₹3.50 | Per 10,000 characters |
| Bulbul v2 | ₹15 | Per 10,000 characters |
| Bulbul v3 | ₹30 | Per 10,000 characters; listed as beta pricing |
| Document Digitization API | ₹0.50 | Per page; listed maximum job size of 10 pages |
Speech requests are listed as rounded up to the nearest second. Enterprise customers may receive custom billing or volume discounts.
Developer onboarding and operational checks
The official welcome pages describe API-key signup, REST access and SDK examples. Sarvam-105B uses an OpenAI-compatible chat-completions format, but authentication headers, endpoints and model identifiers can change.
- Create an account and API key through the Sarvam dashboard.
- Read the current API reference and confirm the base URL, authentication header, endpoint and model name.
- Run a small, representative test set covering each target language, script, accent and audio condition.
- Measure latency, error rates, token or audio usage and output quality before estimating production cost.
- Confirm rate limits, privacy, retention, data location, support, SLA and commercial-use terms in the contract.
Sarvam says rate limits differ across REST, WebSocket, vision and language-model APIs and use continuous token-bucket-style replenishment. Do not assume one requests-per-minute limit applies to every account or endpoint.
API onboarding · Rate limits and credits
Important limitations and open questions
- Language equality is unproven: support lists do not establish equal quality across languages, scripts or accents.
- Sovereignty is not automatically privacy: hosted APIs may still process data under specific retention, logging, cloud and subcontractor terms.
- Open-weight is not the same as fully open source: an Apache 2.0 model license does not by itself provide training data, training code or reproducible infrastructure.
- Benchmarks need independent reproduction: reported scores require checkpoint, prompt, tool and evaluator details.
- Large models cost money to serve: total parameter count does not directly predict quality, speed, factuality or safety.
- Indus remains availability-sensitive: its limited-beta launch means access and features can change.
- Funding is not adoption: venture capital, government compute support, revenue, usage and production customers are different measures.
Sarvam’s June 2026 Series B announcement said the company had a $234 million first close toward a planned $300 million round at a $1.5 billion post-money valuation, with HCLTech and Bessemer Venture Partners investing alongside existing backers. Sarvam also reported 10 million daily API calls and more than half a million hours of speech transcription per month; those are company claims, not independently audited market-share figures.
Series B announcement · Series A announcement
Who should consider Sarvam?
- Developers: evaluate the API when Indian-language text, speech, translation or document processing is a core requirement.
- Enterprises and governments: run a controlled pilot using representative data, then negotiate deployment, retention, support and compliance terms.
- Consumers: try Indus if access is available, but do not assume beta availability or feature parity with established assistants.
- Technical teams considering self-hosting: verify the exact open-weight checkpoint, hardware requirement, license, acceptable-use rules and operating cost before committing.
Bottom line
Sarvam’s importance is not that it has already replaced ChatGPT or Gemini. It is building an Indian-language AI stack—from models and speech systems to translation, document tools, APIs, deployment infrastructure and a consumer interface—at a time when language access and control of critical AI systems are strategic issues. For Indian-language software and public-service workflows, Sarvam is a credible platform to test. For general-purpose consumer use, its fit depends on the maturity and availability of Indus and on whether its local advantages outweigh the broader ecosystems of global competitors.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
Recommended Free Tools
