What is Sarvam-105B?
Sarvam-105B is Sarvam AI's flagship, announced at the India AI Impact Summit in February 2026 with weights on Hugging Face and AIKosh in early March. It is a mixture-of-experts model with about 106 billion parameters, 10.3 billion active per token, and a 128K-token context, pretrained from scratch in India. Sarvam focused training on Indian languages and says it leads its size class across 22 of them, from Hindi, Bengali and Tamil to Santali and Dogri. It reads and writes text only.
The weights are under Apache 2.0, and a smaller Sarvam-30B sits alongside it. Sarvam upgraded the model at its Epoch conference on 30 July, and its 3 November Prism decision model is a LoRA fine-tune of it. A trillion-parameter successor is announced but not released. On Hugging Face it has 324 likes and 5,832 downloads in the last 30 days.
Sarvam's API charges ₹15 per million input tokens, ₹5 for cached input and ₹60 for output, and new accounts get ₹100 of free credit.
How much does Sarvam-105B cost?
| Plan | Price | What you get |
|---|---|---|
| Weights | Free | Apache 2.0 |
| Sarvam API | ₹15 in / ₹60 out per 1M tokens | ₹5 cached input; ₹100 free credit |
Free weights under Apache 2.0. Sarvam API: ₹15 input / ₹5 cached / ₹60 output per 1M tokens; ₹100 free credit on sign-up.
Is Sarvam-105B worth it?
Sarvam-105B is a language model for India first and a general model second. Its case rests on 22 Indian languages and a price in rupees, not on English benchmarks. At launch Artificial Analysis gave it 18 on its Intelligence Index, behind GLM-4.5-Air (23), Mistral Small 4 (27) and gpt-oss-120b (33), all of which use more active parameters.
It does better on agent work than the headline suggests. Artificial Analysis gave it 25 on its Agentic Index, ahead of GLM-4.5-Air's 21, and Sarvam reports 49.5 on BrowseComp and 68.3 on τ² Bench, above gpt-oss-120b's 65.8.
The catch: coding and accuracy. Artificial Analysis measured 1.5% on Terminal-Bench Hard against 20.5% for GLM-4.5-Air, and an Omniscience score of -60, meaning it answers when it should admit it does not know. Sarvam's own SWE-bench Verified figure is 45.0, below every model in its comparison table. The context stops at 128K tokens, and vLLM support still needs a patch or fork.
Our take: Sarvam-105B earns its place behind an Indian-language customer service line, a government form assistant or a voice agent, where fluency in Marathi or Odia matters more than solving a terminal task. At ₹60 per million output tokens it is cheap enough to try first and keep a stronger English model for the hard cases.
Where it earns its place
An Indian bank's customer-support team
Answer customer chats in Hindi, Tamil and Bengali
Call Sarvam-105B through Sarvam's API with retrieval over policy documents. At ₹15/₹60 per million tokens, 100,000 chats of 3,000 input and 500 output tokens cost about ₹7,500.
A state government digital services unit
Run a citizen assistant on government-controlled servers
Self-host the Apache 2.0 weights from Hugging Face or AIKosh with SGLang. Keep a human or retrieval check on facts, since Artificial Analysis found a high hallucination rate.
Before you commit
- Setup
- Sign up for Sarvam's API for ₹100 of free credit, or download the weights and serve with SGLang; vLLM needs Sarvam's fork or the hot-patch script in the repository.
- Integrations
- Sarvam API, which also serves GLM 5.3, Gemma 4 and DeepSeek V4.1 Flash; Hugging Face Transformers, SGLang and patched vLLM.
- Data and privacy
- Self-hosting keeps data in-house. Sarvam's pricing page shows ISO and SOC 2 Type I badges; retention and training terms for API traffic are not on it.
- Cost at scale
- ₹15 input, ₹5 cached and ₹60 output per million tokens; enterprise plans add higher rate limits and custom throughput.
- Limits
- 128K-token context; Sarvam's benchmarks were run at a 65,536-token limit. Rate limits are not published for standard plans.
What’s good
- Trained for 22 Indian languages
- Apache 2.0 licence, pretrained from scratch in India
- ₹15/₹60 per million tokens with ₹100 of free credit
- Agentic Index 25 at launch, ahead of larger GLM-4.5-Air
What’s not
- 1.5% on Terminal-Bench Hard in Artificial Analysis's tests
- High hallucination rate: Omniscience score of -60
- 128K-token context
- vLLM needs a fork or hot-patch
Best for
Indian businesses and public bodies that need an open, affordable model fluent in Indian languages.
Sarvam-105B alternatives
Sarvam-105B: common questions
- Is Sarvam-105B worth it?
- The open model to reach for in Hindi, Tamil or Bengali, as long as nobody asks it to write code or check its own facts. We rate it 3.6 out of 5.
- How much does Sarvam-105B cost?
- Free weights under Apache 2.0. Sarvam API: ₹15 input / ₹5 cached / ₹60 output per 1M tokens; ₹100 free credit on sign-up.
- What are the downsides of Sarvam-105B?
- 1.5% on Terminal-Bench Hard in Artificial Analysis's tests; high hallucination rate: Omniscience score of -60; 128K-token context; vLLM needs a fork or hot-patch.
- Who is Sarvam-105B best for?
- Indian businesses and public bodies that need an open, affordable model fluent in Indian languages.
- How much does Sarvam-105B cost for a team or in production?
- ₹15 input, ₹5 cached and ₹60 output per million tokens; enterprise plans add higher rate limits and custom throughput.
- Is Sarvam-105B safe for company data?
- Self-hosting keeps data in-house. Sarvam's pricing page shows ISO and SOC 2 Type I badges; retention and training terms for API traffic are not on it.
- What are the best Sarvam-105B alternatives?
- The closest alternatives we have reviewed are gpt-oss-120b, Mistral Small 4, Qwen3.8-27B and Command A+.
Answers checked Nov 7, 2026.