Meet changing demand
Serving capacity scales automatically with your traffic, with dedicated GPUs ready for sustained volume.
Inference that scales with demand
Sarvam runs the serving infrastructure behind your applications so capacity, reliability and integration stay out of your way.
Keep production running
Active fleet monitoring and staged rollouts catch faults early and roll back changes before they hit traffic.
Fit your existing tools
Use your existing tools and frameworks, and switch between models with minimal changes to your application.
Speed and reliability
on one serving fleet.
Speed
Faster inference at every stage of the response.
Earlier first tokens and faster generation mean less waiting inside your application.
Reliability
Reliability backed by an SLA
Keep critical applications available with 99.95% uptime guaranteed by an SLA.
SLA-backed
99.95%
Uptime guaranteed
Backed by contractual uptime commitments for enterprise deployments.
-
SLA-driven scaling
Workers scale automatically to hold your latency targets.
-
Continuous monitoring
Active health checks detect and isolate faults across the fleet.
-
Staged upgrades
Rollouts happen in stages with automated checks and instant rollback.
Built for real-world workloads
From multi-step code edits to live voice streams, run every workload on one API.
Coding agents
Read repo context, reason through bugs and run tools to ship working patches.
Tool-calling agents
Connect models to your APIs and business systems with structured JSON outputs.
Million-token context
Pass whole documents, instructions and chat history together in a 1M-token window.
Document vision
Read pages, tables and charts alongside text to pull out structured data.
Multi-step reasoning
Tackle complex analysis and planning with adjustable reasoning effort.
Real-time streaming
Stream tokens straight into chat UIs and voice pipelines as they generate.
Leading models. Ready to use
Explore models for reasoning, coding, vision and Indian languages.
-
GLM 5.3
- Z.ai
- Reasoning and coding
- 1M context
Large-scale reasoning model built for complex software engineering and long-horizon agent tasks.
-
GLM 5.3 Flash
- Z.ai
- Reasoning and coding
- 1M context
Faster open-weight reasoning model for agents and image understanding.
-
DeepSeek V4.1 Flash 1017
- DeepSeek
- Reasoning and coding
- 1M context
General-purpose open-weight model for long-document analysis and agentic workflows.
-
Gemma 4 31B
- Google DeepMind
- Vision
- 256K context
31B-parameter instruction-tuned model for text and image understanding, with tool calling and optional reasoning.
-
Sarvam 105B
- Sarvam
- Indian languages
- 128K context
Flagship 105B Mixture-of-Experts model trained from scratch for Indian-language understanding, complex reasoning, coding and agentic workflows.
-
Sarvam 105B Conversations
- Sarvam
- Indian languages
- 32K context
Post-trained variant of Sarvam 105B for real-time dialogue, voice agents and multi-turn conversations across Indian languages.
Serverless inference. Ready to scale
Run leading models through an OpenAI-compatible API. Sarvam manages the serving infrastructure in India.
pip install --upgrade sarvamaiV2 · Sarvam SDK · Streamingimport osfrom sarvamai import SarvamAIclient = SarvamAI(api_subscription_key=os.environ["SARVAM_API_KEY"])stream = client.chat.completions_v2(model="glm5.3",messages=[{"role": "user", "content": "Explain serverless inference briefly."}],max_tokens=8000,stream=True,)for chunk in stream:if not chunk.choices:continuetext = chunk.choices[0].delta.contentif text:print(text, end="", flush=True)
V2 Chat Completions. Max reasoning is the default. Reasoning tokens count toward the 8,000-token output budget.
From first request to production traffic
Access supported models on Sarvam’s GPUs in India, with serving capacity that adapts to your traffic.
- Model evaluation and prototyping
- Production applications
- Variable or unpredictable traffic
Enterprise-grade security.
Compliance and privacy built in.
Independently audited
ISO 27001:2022 certified and SOC 2 Type I attested by independent assessors for information security and operational controls.
Your inference stays in India
Prompts run on Sarvam GPUs in India and are never sent to overseas model providers.
Enterprise-grade security
Every request is encrypted, isolated to your account and scoped to your team.
- Encrypted in transit and at rest
- Workloads isolated by account
- Role-based access for keys and teams
Privacy on your terms
Enable zero data retention on request, and set clear rules for logs.
- Prompts and responses Zero retention Never stored or used to train models On request
- Operational logs Custom retention
Questions before you go live
Answers for teams evaluating models, moving applications and planning production.
Sarvam supports OpenAI-compatible Chat Completions and Responses APIs. Update your endpoint, key and model, then validate the parameters, tool calls and outputs your application uses. Read the API guide
Leading open models perform competitively on coding, reasoning and other published evaluations. The right choice depends on your workload. Validate answer quality, tool use and response times on representative requests. Compare published scores on Artificial Analysis.
Model inference runs on Sarvam-operated GPUs in India. Sarvam hosts the weights; requests are not forwarded to the original model developer’s overseas API. External tools connected by your application have their own processing locations and data policies.
Zero data retention is available on request. Under an agreed ZDR arrangement, prompts and responses are not retained. Operational logs are scoped separately in the terms for your deployment.
No. Sarvam does not use customer data to train models for other customers. The original model developer does not receive your inference requests. Your data-processing agreement defines the permitted use and handling of your data.
Sarvam manages serving capacity as traffic changes. Share your expected concurrency, prompt lengths, peak traffic and response-time targets so the team can size the deployment. Available capacity and rate limits depend on your model, deployment and account.
Yes. Select supported models through the API’s model field. Context limits, input types and tool behaviour vary, so validate your application when changing models. You can use all models listed in the catalogue with a single API key by changing the model name.
Sarvam’s engineers can support evaluation, integration and migration. For production, discuss the capacity plan, availability targets, support response times and escalation process with the team. Service commitments are agreed for your enterprise deployment. Discuss SLA
Request Sarvam’s ISO 27001:2022 certificate, SOC 2 Type I report, penetration-test summary and data-processing documentation through the Trust Center. Some documents are shared under NDA. Your security team can review their scope against the planned deployment. Visit the Trust Center
Test it on your own traffic. Our engineers help you compare quality, speed and cost before you move.
Test it on your own traffic.
Our engineers help you compare quality, speed and cost before you move.