Sarvam-M

What it is

Sarvam-M is an open-weights large language model released by Sarvam AI. It has 24 billion parameters and was not trained from scratch: it is a post-trained version of Mistral Small, an existing open model distributed under the Apache-2.0 licence — which is where the "M" in the name comes from.^[1] The released weights are Apache-2.0 as well, and the model covers eleven languages, including Hindi, Bengali, Tamil and Telugu.^[2]

Post-training means everything a lab does to a base model after its first round of training on raw text: teaching it to follow instructions, shaping its tone, and pushing its accuracy up on the tasks that matter. Because the base already has general language ability, this stage is far cheaper than pretraining and can be aimed at a particular audience.

Where it came from

Sarvam AI published the technical blog describing the model on 23 May 2025.^[1] The weights were uploaded three days earlier, on 20 May 2025.^[2] Post-training an existing 24-billion-parameter open model rather than building a base model from scratch keeps the compute budget on behaviour and language coverage instead of on raw pretraining.

How it was trained

Sarvam describes three parts: supervised fine-tuning, reinforcement learning with verifiable rewards, and inference optimisation.^[1]

Supervised fine-tuning is the direct part. The model is shown many pairings of an instruction with a good answer and learns to imitate the format and style of those answers.

Reinforcement learning with verifiable rewards is used where correctness can be checked automatically — a maths problem with one right answer, or code that either runs or does not. The model generates attempts, the checkable ones earn a reward, and training nudges the parameters toward whatever earned it. The algorithm used here was GRPO, short for group relative policy optimisation, which compares a group of sampled answers to one another instead of training a separate model to score each one.

Inference optimisation is the third part: changes made so the model is cheaper or faster to serve, rather than sharper on a benchmark.^[1]

Sarvam also says it debiased completions carrying a political slant and re-biased the model toward culturally relevant outputs.^[1] In plain terms, the post-training data was edited in two directions: answers that leaned one way politically were removed, and the model was deliberately steered toward material that fits the cultures it is meant to serve.

The model runs in two modes, think and non-think, both trained through a single curriculum.^[1] In think mode it writes out intermediate reasoning before answering; in non-think mode it replies straight away. Coming from one curriculum, the same weights can do both, and the caller picks per request.

Results

Against the base model, Sarvam reports a 20% average gain on Indian language benchmarks.^[1][2] Maths benchmarks rose 21.6% and programming benchmarks 17.6%.^[1][2] The largest single jump is on a GSM-8K benchmark in romanised Indian languages, up 86%.^[1] GSM-8K is a set of grade-school maths word problems; "romanised" means Indian languages written in Latin letters, the way they are often typed in chat. That is a very different input from the standard script, and the size of the gain suggests the base model found it hard.

On MMLU-IN, an India-focused version of the MMLU knowledge benchmark, the score moved from 0.64 to 0.79; on MMLU-IN-R it moved from 0.49 to 0.66.^[1]

Limitations

English MMLU slipped by about one point against the base model, which Sarvam lists as a weakness.^[1] The pattern — large gains in the target languages, a small cost in English — is the trade post-training makes: capacity is steered toward Indic data and behaviour, and something is given up elsewhere.

Where it sits

An Apache-2.0 checkpoint covering eleven Indian languages, published alongside an account of how it was trained, is something other teams can fine-tune, self-host or build on.^[2] The reported gap over its own base model on maths, code and Indic benchmarks is the useful part of the claim: it measures what the post-training stage added, not what a 24-billion-parameter model can do in general.