Intel / article

Sarvam 105b Sarvam 30b Everything You Need To Know

Source: artificialanalysis.ai
Author: Artificial Analysis articles
Date: Published Apr 2, 2026

Context

SarvamAI's Sarvam 105B and Sarvam 30B are, in Artificial Analysis's description, India's largest models pre-trained from scratch (artificialanalysis.ai). from scratch means training the model's weights from the beginning rather than adapting an existing model: Sarvam's previous model, Sarvam M, was based on Mistral AI's Mistral Small. Both new models were announced at the India AI Impact Summit 2026, trained entirely in India on compute provided under the IndiaAI Mission, and released as open weights under the Apache 2.0 license on HuggingFace and AIKosh. They score 18 and 12 on the Artificial Analysis Intelligence Index, up from 8 for Sarvam M.

Both are models, which send each token through only part of the network, so far fewer parameters are active per token than the model holds in total. Sarvam 105B has 106B total parameters, about 10B active per token, and a 128K-token ; Sarvam 30B has 32B total, about 2.4B active, and a 65K window. Against open-weights reasoning models of similar size, both trail. Sarvam 105B is behind GLM-4.5-Air (23), INTELLECT-3 (22), Mistral Small 4 (27) and gpt-oss-120B (High, 33), all of which activate more parameters per token. Sarvam 30B is behind GLM-4.7-Flash (30), Nemotron Cascade 2 30B A3B (28), Qwen3 30B A3B 2507 (22) and Qwen3 32B (17), while activating fewer parameters than any of them.

The results are uneven across tasks. Sarvam 105B's agentic index of 25 puts it ahead of INTELLECT-3 (20) and GLM-4.5-Air (21) despite trailing both overall, and its GDPval score of 773 edges GLM-4.5-Air's 665. Both models are weak on TerminalBench Hard, a test of agentic coding and terminal use, at 1.5% for Sarvam 105B and 2.3% for Sarvam 30B against 20.5% for GLM-4.5-Air and 9.1% for INTELLECT-3. On AA-Omniscience, which docks points for wrong answers, they score -60 and -72. Artificial Analysis attributes this to rates that are high relative to accuracy and to attempting far more questions rather than abstaining. Both models are text-only and currently free on Sarvam's own API.

Key takeaways · AI-distilled

Sarvam 105B and 30B are mixture-of-experts models pre-trained from scratch in India on IndiaAI Mission compute under Apache 2.0. The 105B activates about 10B parameters with a 128K window; the 30B about 2.4B with a 65K window.
Sarvam 105B's relative strength is agentic work: its agentic index of 25 beats INTELLECT-3 (20) and GLM-4.5-Air (21), and its GDPval index of 773 edges GLM-4.5-Air's 665, despite lower overall intelligence.
On TerminalBench Hard, Sarvam 105B scored 1.5% and Sarvam 30B 2.3%, against 20.5% for GLM-4.5-Air. Their AA-Omniscience scores of -60 and -72 reflect attempting far more answers instead of abstaining.
Both are a large step up from Sarvam M, which scored 8 and was built on Mistral Small rather than pre-trained from scratch. Sarvam also announced Saaras v3 speech-to-text and Bulbul v3 text-to-speech for Indic languages.

Terms in this piece · Glossary

pretraining — The first, biggest phase of building a model: training it on enormous amounts of text so it learns language, facts, and reasoning in general.
mixture-of-experts — A model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
hallucination — When a model states something false with full confidence — inventing facts, citations, or APIs that don't exist.

Read the source artificialanalysis.ai