Sarvam AI: Sarvam-105B

Overview

Sarvam AI's flagship reasoning model, trained from scratch: a Mixture-of-Experts with 105B total and 10.3B active parameters and Multi-head Latent Attention, built for Indian languages and English, coding, maths and agentic tool use, released as open weights under Apache 2.0.

Provider

Sarvam AI

Modality

Text

Input

text

Output

text

Parameters

105BProvider's model page, checked October 9, 2026.

Weights

Open weights

Supported parameters

tools, tool_choice, structured_outputs, response_format, reasoning, temperature, top_p, max_tokens, stop, seed, frequency_penalty, presence_penaltyProvider's model page, checked October 9, 2026.

Distillable

Conditional (licence terms)Apache 2.0 weights: outputs of a copy you host may train other models. Sarvam's terms (section 10.5) forbid using output from Sarvam's API to train or improve any AI model without its written permission.Licence or terms, checked October 9, 2026.

Tags

open-weights, reasoning, multilingual, indian-languages, coding, agents, tool-use

Docs

Provider documentation

Data policy

Prompts go to

Sarvam AI through its API: the managed environment runs in Microsoft Azure's Central India region and LLM inference on GPU infrastructure in India run by Yotta and NxtGen. A copy of the open weights you host sends nothing to Sarvam.

Retention

Sarvam keeps Model API data until the workspace Owner sets a retention period, and No retention (0 days) is available for Model APIs. Sarvam's terms let it train its own models on inputs and outputs as its Privacy Policy and the law allow, with consent where required, and its Trust Center says one customer's data never trains models for another customer.

Zero data retention

Available

Region

India: Sarvam's managed production environment runs in Microsoft Azure's Central India region.

Policy

Sarvam AI data policy

Summarised from the provider's published terms. The provider's own policy is what applies.

Where you can run it

How Sarvam-105B can be deployed, bought and reached through Safeguard.

Deployment

Public cloudPrivate cloudOn-premAir-gapped

Procurement

Not published yet

Inference regions

🇮🇳IndiaCloud provider's availability page, checked October 9, 2026.

On-prem and air-gapped deployments run wherever you install them.

Pricing

Notes

Sarvam lists INR 29.28 input, INR 10.98 cached input and INR 73.20 output per 1M tokens. Prices are in Indian rupees only, so no US dollar price is shown.

Source

Sarvam AI pricing , last verified October 9, 2026

No markup on model usage: you pay the provider's list price.

API

OpenAI-compatible. Your API key and access to this model are set up for your workspace when you choose Use with Safeguard.

from openai import OpenAI

client = OpenAI(
    base_url="https://api.safeguard.sh/gpt/v1",
    api_key="SAFEGUARD_API_KEY",
)

response = client.chat.completions.create(
    model="sarvam/sarvam-105b",
    messages=[{"role": "user", "content": "Hello"}],
)
print(response.choices[0].message.content)

Base URL https://api.safeguard.sh/gpt/v1, model sarvam/sarvam-105b.