Sarvam AI

Overview

Sarvam AI is an Indian company whose stated goal is to build a sovereign AI stack in India — model development the company presents as its own, rather than assembled entirely from outside suppliers. On 23 May 2025 it published the first entry in a planned series of technical blogs about that work; the entry introduced a model called Sarvam-M^[1].

Where Sarvam-M comes from

Sarvam-M is a post-trained version of Mistral Small, a 24-billion-parameter model released under the Apache-2.0 licence, and the M in the name stands for Mistral^[1]. Post-training means starting from a base model that has already absorbed general language patterns from large amounts of text and continuing to train it against narrower objectives, rather than training from scratch. The weights were uploaded on 20 May 2025, three days before the blog appeared, and a playground and APIs shipped at the same time^[1][2].

How it was trained

Sarvam describes three parts to the work: supervised fine-tuning, reinforcement learning with verifiable rewards, and inference optimisation^[1]. Supervised fine-tuning trains a model on example inputs paired with desired outputs. Reinforcement learning with verifiable rewards goes a step further: the model produces candidate answers, and a program checks whether each one is correct — for a mathematics problem, whether the final result matches; for code, whether it passes tests — and the model is pushed toward the answers that check out. Inference optimisation is separate engineering work aimed at making the trained model cheaper or faster to serve.

The model runs in think and non-think modes trained through a single curriculum^[1]. Think mode means the model writes out intermediate reasoning before answering; non-think mode answers directly. Training both through one curriculum means the two behaviours were taught together rather than as two separate models.

For the reinforcement-learning step Sarvam used GRPO^[1]. GRPO (group relative policy optimisation) samples a group of answers to the same prompt and compares them with one another, using their relative scores to decide which answers to reinforce, rather than relying on a separate model to estimate how good each answer is.

The company also says it debiased completions with a political slant and re-biased them toward culturally relevant outputs^[1]. In plain terms, training data and outputs were adjusted so that answers would not carry a political leaning the company considered inappropriate, and so that they would fit the cultural context of Indian users.

Results

Sarvam reports a 20 percent average gain on Indian-language benchmarks over the base model, with mathematics up 21.6 percent and programming up 17.6 percent^[1][2]. MMLU-IN rose from 0.64 to 0.79 and MMLU-IN-R from 0.49 to 0.66, while English MMLU slipped about one point^[1]. MMLU is a standard set of knowledge questions; the -IN versions are adapted to Indian languages and context, so gains there are the specific outcome the post-training was aimed at. The small English drop suggests the gain came with a trade-off.

Sarvam says the model beats Llama-4 Scout on most benchmarks and is comparable to Llama-3.3 70B and Gemma 3 27B^[1]. Sarvam-M carries 24 billion parameters, so those comparisons set it against considerably larger models.

Openness and language coverage

The weights are Apache-2.0 and cover eleven languages including Hindi, Bengali, Tamil and Telugu^[2]. Apache-2.0 is a permissive open-source licence that allows commercial use and modification, provided copyright and licence notices are kept.

How it fits

Sarvam-M is the first entry in the planned series on building a sovereign AI stack in India^[1], so it reads as a public starting point for the company: a working model, a set of comparable scores and open weights, with the rest of the stack left for later posts.