Sarvam 105B: India's open reasoning model, how to try it

Sarvam 105B is an open reasoning model from Sarvam AI, a Bengaluru company, trained from scratch in India. It is a Mixture-of-Experts model with about 106 billion total parameters and 10.3 billion active parameters, released under the Apache 2.0 licence.

We look at what the model is, what Sarvam claims about it, how you can try it today, and what a student can research with it.

What is Sarvam 105B?

Sarvam 105B is a large language model built for reasoning, coding and Indian languages. Sarvam open-sourced it together with a smaller model, Sarvam 30B, in a blog post titled "Open-Sourcing Sarvam 30B and 105B", published on 6 March 2026.

Sarvam says the models were "trained from scratch on large-scale, high-quality datasets curated in-house." The blog adds: "Training was conducted entirely in India on compute provided under the IndiaAI mission." The 105B model was trained on 12 trillion tokens.

How big is Sarvam 105B, and how is it built?

Sarvam 105B uses a Mixture-of-Experts (MoE) design. An MoE model has many small expert networks, and a router picks a few of them for each token. So only part of the model runs at a time. If MoE is new to you, our post on small language models explains why active parameters matter for cost.

The facts from the Hugging Face model card:

• Size: about 106B total parameters, 10.3B active per token.
• Experts: 128 routed experts, top-8 routing, plus 1 shared expert.
• Attention: MLA-style attention.
• Context: 128K tokens, using YaRN scaling.
• Licence: Apache 2.0, so you can use it for research and commercial work.

Which languages does Sarvam 105B support?

Sarvam 105B is aimed at English and Indian languages. The model card says it delivers "state-of-the-art performance across 22 Indian languages" for its size. The Sarvam blog says the tokenizer is "optimized for efficient tokenization across all 22 scheduled Indian languages."

The training data is not split evenly. The blog says: "The multilingual corpus allocates a substantial portion of the training budget to the 10 most-spoken Indian languages." So expect better results in Hindi or Bengali than in a smaller scheduled language. Test your own language before you rely on it.

Does Sarvam 105B have a thinking mode?

Yes. Sarvam 105B is a reasoning model that writes its thinking inside `<think>` tags before the final answer. In the Hugging Face code examples, thinking is turned on with `enable_thinking=True` in the chat template. Neither the blog nor the card documents a thinking budget setting. To learn how this kind of model works, read our guide to reasoning models.

How good is Sarvam 105B? Benchmarks as Sarvam reports them

The numbers below are Sarvam's own results from the model card. We have not run them ourselves, and independent tests may differ.

• MMLU: 90.6
• MMLU Pro: 81.7
• GPQA Diamond: 78.7
• Math500: 98.6
• AIME 25: 88.3 (96.7 with tools)
• LiveCodeBench v6: 71.7
• SWE-Bench Verified: 45.0
• BrowseComp: 49.5

Sarvam compares these with GLM-4.5-Air, GPT-OSS-120B and Qwen3-Next-80B-A3B-Thinking. On SWE-Bench Verified, Sarvam 105B scores below all three, so coding agents are not its strongest area.

For Indian languages, Sarvam reports that 105B "significantly outperforms all models we evaluated." Its test used 110 English prompts translated into 22 Indian languages, in native and romanised script. Gemini 3 judged the answers, and Sarvam reports that 105B won about 90% of pairwise comparisons. An LLM judge is a useful tool, but it is not the same as native speakers rating answers.

How to try Sarvam 105B

You can try Sarvam 105B in three ways, from easiest to hardest:

1. 1Use the Indus app. Sarvam says the 105B model powers Indus, its chat app, at indus.sarvam.ai. This needs no setup.
2. 2Use the API. Sarvam says both models are on its API dashboard at dashboard.sarvam.ai. Check the current pricing there before you build on it.
3. 3Download the weights. The model is on Hugging Face and on AI Kosh. The card has code for Transformers, SGLang and vLLM.

Running the full model yourself needs several data-centre GPUs. The card's SGLang example uses 4 GPUs and its vLLM example uses 8. vLLM support was not merged at the time of the card, so Sarvam gives a custom fork and a patch script. The card also lists quantised versions for llama.cpp, Ollama and LM Studio, which are easier on smaller machines. For a student laptop, Sarvam 30B or the app is the practical choice.

Why do Indian-language AI models matter?

Indian-language models matter because most Indians do not use English as their first language. Many popular models are trained mostly on English text. They often split Indian scripts into many tokens, which makes them slower and costlier to run in Hindi or Tamil than in English.

A tokenizer built for Indian scripts, and training data chosen for Indian languages, can help with both cost and quality. An open licence also means Indian universities, startups and government projects can study and adapt the model, instead of only calling a foreign API.

Research ideas for students using Sarvam 105B

An open model with Indian-language claims gives students many testable questions. Here are some you could start with:

• Low-resource test: build a small question set in a less-spoken scheduled language, such as Santali or Bodo, and compare it with Hindi.
• Judge check: rate 50 answers with native speakers and compare with an LLM judge. Do they agree?
• Token cost: count tokens for the same sentence in English and five Indian languages, across several tokenizers.
• Code-mixed text: test Hinglish or romanised Bengali, which people actually type on WhatsApp.
• Thinking length: does the model think longer in Indian languages than in English for the same maths problem?
• Local use cases: test answers on Indian law, farming or exam syllabi, and log factual errors.

Keep each project small and measurable. Write down the prompts, the model version and the settings so others can repeat it. If you want support from government-backed programmes, see our guide to the IndiaAI Fellowship 2026-27. To find a team and mentors for a project like this, join APJ-RAISE. It is free.