SearchBook a call

BULLETIN / Sarvam AI

Sarvam 30B and 105B: review

Sarvam's open models cover all 22 scheduled Indian languages under Apache 2.0, but no independent tests exist yet and agentic coding scores are modest.

The first open model family to try for Indian languages, pending independent evaluation.

Maker
Sarvam AI
Released
6 Mar 2026
Licence
Apache-2.0
Access
Open weights (Hugging Face, AI Kosh), Sarvam API
Context
65,536 tokens (Sarvam 30B)
Price
Sarvam 105B: ₹29.28 input / ₹73.20 output per million tokens (INR)

Good at

  • All 22 scheduled Indian languages
  • Small active size for cheap serving (30B)
  • Apache 2.0 licence
  • Maths and general knowledge for its size

Bad at

  • No independent evaluation we could find
  • Short context for the 30B
  • Modest agentic coding scores
  • Diagram-based reasoning

Use it for

  • Indic language chat and support products
  • Teams that need data and training to stay in India
  • Fine-tuning for regional language tasks

Avoid it for

  • Long coding agents
  • Very long documents on the 30B
  • Image input

Yes, if your users speak Indian languages. Sarvam 30B and 105B are open models trained from scratch in India, cover all 22 scheduled Indian languages, and use the Apache 2.0 licence. The caveat is evidence: every score we found comes from Sarvam, and its own scores on the software engineering benchmark SWE-Bench Verified are modest.

First principles

Language skill comes from training data. A model learns a language from the text it is trained on, and most large models see far more English than any Indian language. Training from scratch with all 22 languages as a design goal is the right starting point, but only measurement shows how far it gets.

Tokens are the unit of both cost and context. Models read text as tokens. A tokeniser built mostly from English text tends to split Indic scripts into more tokens per word, which raises the cost of every sentence and fills the context window sooner. Check token counts on your own text, especially on the 30B with its 65,536-token window.

Active parameters set serving cost. In a mixture-of-experts model only some experts run for each token. The 30B runs 2.4B of its 32B parameters per token, so its compute per token is close to a small model’s, though all 32B must still fit in memory.

A vendor’s own score is a claim, not a check. The vendor picks the tests, the settings and which results to publish. Until an independent board runs the same tests on Sarvam and its rivals, compare it on your own data.

What it is

Sarvam AI open-sourced two models on 6 March 2026. Both use a mixture-of-experts design, where only part of the model runs for each token.

  • Sarvam 30B has 32B total parameters and 2.4B active. Its context window is 65,536 tokens.
  • Sarvam 105B uses 128 experts. Sarvam’s post does not state its active parameter count.

Sarvam says it built every stage of training in-house and trained the models entirely in India on compute from the IndiaAI Mission. The 30B saw 16T tokens in pre-training, the 105B 12T. The models cover the 22 scheduled Indian languages (those listed in the Eighth Schedule of the Constitution of India) across 12 scripts, plus English. Weights are on Hugging Face and on the AI Kosh repository.

Where it is good

Indian languages. This is the reason to use Sarvam. Few open models from any vendor name all 22 scheduled languages as a design goal. Among the models in this edition, IBM Granite 4.2 lists none, and Xiaomi’s MiMo card lists only English and Chinese.

Serving cost. With 2.4B active parameters, the 30B is cheap to run for its total size. The model card says it was designed for resource-constrained deployment, and it runs on vLLM, SGLang and Hugging Face Transformers.

Licence and location. Apache 2.0 allows commercial use and fine-tuning. For Indian public-sector or regulated buyers, a model trained in India by an Indian company may simplify procurement.

Vendor-reported scores are solid:

  • Sarvam 105B: 90.6 on Massive Multitask Language Understanding (MMLU), 98.6 on Math500, 71.7 on LiveCodeBench v6.
  • Sarvam 30B: 85.1 on MMLU, 97.0 on Math500, 92.1 on HumanEval.

These are Sarvam’s numbers, not independent results.

Where it is bad

No independent check. We did not find Sarvam in the Artificial Analysis, Arena, Vals or Epoch AI results we read. That makes it hard to compare directly with Gemma 4 or Qwen3.8-27B, which do appear there.

Sarvam’s own numbers disagree in places. For the American Invitational Mathematics Examination (AIME) 25, the launch post gives the 30B 80.0 at pass@1 (first-attempt accuracy), while the model card gives 88.3. We report neither as settled.

Agentic coding. Sarvam reports 34.0 for the 30B and 45.0 for the 105B on SWE-Bench Verified with the SWE-Agent harness. These are vendor-reported. In the comparison table on the 30B model card, GLM 4.7 Flash scores 59.2 and Nemotron-3-Nano-30B 38.8 on the same test. In Sarvam’s own evaluation on the Joint Entrance Examination (JEE), most physics errors came from diagram-based reasoning.

Short context. At 65,536 tokens, the 30B holds far less than the 128K to 1M windows common elsewhere. Long contracts or codebases will need chunking.

Text focus. We found no statement of image input for these models, so on this evidence treat them as text-only models.

Price and access

Model Input per million Cached input Output per million
Sarvam 105B (hosted) ₹29.28 ₹10.98 ₹73.20
Sarvam 30B Free weights Not applicable Not applicable

Prices are in Indian rupees. Sarvam offers free starter credits but does not state the amount.

The wider Indian open-model field is thin. BharatGen’s Param2 17B model covers English, Hindi and 21 other Indian languages, but it uses a non-commercial licence and its card calls it an early checkpoint. Krutrim has shifted its focus to cloud services, according to TechCrunch.

Who should use it

On this evidence Sarvam 30B is the first open model to test for Indian-language chat, support, search and summarisation, especially where data or training provenance must stay in India. Use the 105B through Sarvam’s application programming interface (API) for harder tasks. Run your own evaluation in your target languages before launch, and keep a frontier model as a fallback for long coding or reasoning work.

Sources