Atlas

Models

← All models

Sarvam M

Sarvam AI · 2025-05-23 · 23.6B parameters

Sarvam M is Sarvam AI's 23.6B multilingual hybrid-reasoning model fine-tuned from Mistral-Small 24B with a focus on Indian languages and contexts, supporting 10 Indic languages alongside English. It was released in May 2025 and features a Hybrid Thinking Mode that lets users toggle between standard and extended chain-of-thought reasoning, with particular strength in mathematics, coding, and Indic-language tasks. The model was developed by India-based Sarvam AI as part of their mission to build sovereign AI infrastructure for the Indian language ecosystem.

Benchmark scores

BenchmarkScore
AA-LCR0.0
AA-Omniscience Index-62.3
Artificial Analysis Intelligence Index3.0
Artificial Analysis Omniscience Accuracy14.8
Artificial Analysis Omniscience Hallucination Rate90.5
Capability123.2
CritPt0.0
GPQA Diamond41.6
Humanity's Last Exam (Text-Only)3.3
IFBench (Artificial Analysis)31.8
LiveCodeBench (Artificial Analysis)29.5
MATH-500 (Artificial Analysis source)84.7
MMLU Pro69.6
SciCode (Artificial Analysis)17.8
Terminal-Bench Hard2.3
τ²-bench (Telecom)0.0
Loading Atlas data…