Sarvam 30B and 105B

India's from-scratch open reasoning models: a 105B mixture of experts (10.3B active) and a 30B (2.4B active), 128K context, Apache 2.0.

In short. Sarvam 30B and 105B are two large language models from the Indian company Sarvam AI, built from scratch in India with a focus on Indian languages. They are good at reasoning, coding and tool use, and only a small share of each model works on any one word, which keeps them fast. The weights are free under Apache 2.0, and the test scores come from Sarvam.

Summary

Sarvam 30B and 105B are reasoning models from Sarvam AI, an Indian company, released as open weights in early March 2026. Sarvam says it trained both from scratch on data curated in-house for pre-training, supervised fine-tuning and reinforcement learning, entirely in India on compute from the IndiaAI mission. Both are mixture-of-experts transformers with 128 experts and a context of up to 128K tokens. The 105B has about 10.3B active parameters and Multi-head Latent Attention, and the 30B has about 2.4B active parameters and grouped-query attention. Sarvam says it pre-trained the 30B on 16 trillion tokens and the 105B on 12 trillion, with a large share of the multilingual data going to the ten most spoken Indian languages.

In Sarvam's own tables the 105B scores 98.6 on Math500, 71.7 on LiveCodeBench v6, 81.7 on MMLU Pro, 78.7 on GPQA Diamond and 68.3 on τ² Bench, against GLM-4.5-Air, GPT-OSS-120B and Qwen3-Next-80B-A3B-Thinking. It trails GPT-OSS-120B on GPQA Diamond, AIME 25 and HMMT, and its 45.0 on SWE-bench Verified is below GLM-4.5-Air's 57.6. The 30B scores 97.0 on Math500, 70.0 on LiveCodeBench v6 and 80.0 on MMLU Pro, but on the card's MILU benchmark of Indian-language knowledge it scores 76.8 against 82.6 for Qwen3-30B-Thinking-2507. Sarvam also says both models lead on its Indian-language tests across 22 languages. The comparison models were available in early 2026, not the newer open models in this catalog, and we have not reproduced any of these numbers.

The weights use Apache 2.0 and are on Hugging Face and AI Kosh. The main repositories hold float32 weights, 424 GB for the 105B and 129 GB for the 30B, so most people will use the FP8 builds of 113 GB and 39 GB or the Q4_K_M GGUF files of 64 GB and 20 GB. Sarvam runs the 30B in its Samvaad conversational agent platform and the 105B in its Indus assistant. In the 30 days to 11 November 2026, Hugging Face counted about 5,300 downloads for the 105B and 3,900 for the 30B.

Good for

  • Assistants and voice agents for Indian languages, with the 30B designed for low-latency conversation and tool calls
  • Reasoning, coding and search agents that need a 105B model with only 10.3B active parameters
  • Organizations that want a model trained from scratch in India under Apache 2.0 for sovereignty reasons

Watch out for

  • Every benchmark is Sarvam's own, and the tables compare against models available in early 2026 rather than the newest open ones
  • The main repositories hold float32 weights of 424 GB and 129 GB, so use the FP8 or GGUF builds unless you need full precision
  • The 105B trails GPT-OSS-120B on several math and reasoning tests in Sarvam's table, and its SWE-bench Verified score is below GLM-4.5-Air's
  • On the card's MILU benchmark the 30B scores below Qwen3-30B-Thinking-2507, so check Indian-language quality on your own tasks

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Estimates for the Sarvam-30B Q4_K_M (GGUF) variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.

Variants Parameters BF16 FP8 / INT8 4-bit (Q4) Fits on
Sarvam-30B FP32 32.1B 2.4B active 74 GB 37 GB 22 GB Gaming GPU
Sarvam-30B FP8 FP8 32.1B 2.4B active 74 GB 37 GB 22 GB Gaming GPU
Sarvam-30B Q4_K_M (GGUF) INT4 32.1B 2.4B active 74 GB 37 GB 22 GB Gaming GPU
Sarvam-105B FP32 106B 10.3B active 244 GB 122 GB 73 GB Workstation
Sarvam-105B FP8 FP8 106B 10.3B active 244 GB 122 GB 73 GB Workstation
Sarvam-105B Q4_K_M (GGUF) INT4 106B 10.3B active 244 GB 122 GB 71 GB Workstation

Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.

VRAM calculator

The license in plain words

One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.

More about Apache 2.0

Frequently asked questions

What is Sarvam 30B and 105B?
Sarvam 30B and 105B are two large language models from the Indian company Sarvam AI, built from scratch in India with a focus on Indian languages. They are good at reasoning, coding and tool use, and only a small share of each model works on any one word, which keeps them fast. The weights are free under Apache 2.0, and the test scores come from Sarvam.
Can I use Sarvam 30B and 105B commercially?
Yes. Sarvam 30B and 105B is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is Sarvam 30B and 105B open source?
Sarvam 30B and 105B is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
Can I run Sarvam 30B and 105B on my own computer?
Yes, on a good graphics card. The smallest variant, Sarvam-30B Q4_K_M (GGUF), needs about 22 GB of memory at 4-bit precision, which suits a 24 to 32 GB gaming GPU or a Mac with 32 GB or more.
When was Sarvam 30B and 105B released?
Sarvam 30B and 105B was released by Sarvam AI on March 6, 2026.
How long a context window does Sarvam 30B and 105B have?
Sarvam 30B and 105B supports a context window of up to 128K tokens.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
Sarvam Prism Sarvam AI 105B 10.3B active 128K Apache 2.0 (Permissive) 120 GB Big-memory desktop
GLM-5.3-Flash Z.ai (Zhipu AI) 320B 18B active 1M MIT (Permissive) 221 GB Big-memory desktop
gpt-oss OpenAI 21B – 117B 5.1B active 128K Apache 2.0 (Permissive) 15 GB Laptop or phone
Qwen3.6 Alibaba Qwen 27.8B – 36B 3B active 256K Apache 2.0 (Permissive) 19 GB Gaming GPU