Sarvam AI open-sources an Indic reasoning model built on Mistral Small

Background

Sarvam AI presented Sarvam-M as the first entry in a planned series of technical blog posts on building a sovereign AI stack in India ^[1]. The phrase sovereign AI stack generally refers to a country or region owning the pieces it needs to build and run AI systems itself — models, data, serving infrastructure — rather than depending entirely on outside providers. Work of that kind is not only about training a model; it also means writing down how the model was built and served so that local teams can reuse the methods.

What was released

Sarvam AI announced the model on 23 May 2025 ^[1]. Three things went out together with the announcement: the model weights, a playground for trying the model directly, and APIs for developers who want to call it from their own software ^[1]. The weights are published under Apache-2.0, a licence that allows commercial use ^[2]. For anyone outside the company, open weights mean the model can be downloaded and run on their own machines, and can be further trained on their own data; the licence is what decides whether that downstream use is permitted at all.

Base model and post-training

The team chose Mistral Small as the base model for two stated reasons: it has 24B parameters and it is released under Apache-2.0 ^[1]. Parameter count and licence are the two practical questions that come first when picking a foundation to build on. A 24B model is small enough to serve without an enormous cluster, and a permissively licensed base leaves room to publish the derived model on permissive terms as well.

The blog describes the post-training stage in some detail, in particular how training prompts were curated: examples were scored for quality and for hardness, then clustered and sampled ^[1]. Scoring for quality is a way of keeping well-formed, reliable examples and dropping weak ones; scoring for hardness keeps the cases that a model is likely to fail, which are the ones that teach it something during training. Clustering groups similar prompts together, and sampling from those clusters spreads the training signal across topics instead of letting one kind of prompt dominate.

Benchmark results

According to the blog, Sarvam-M beats Llama-4 Scout on most benchmarks and is comparable to Llama-3.3 70B and Gemma 3 27B ^[1]. The comparison is worth reading alongside the base model: Sarvam-M is built on a 24B base, and one of the models it is described as comparable to is a 70B model, roughly three times larger in parameter count.

Why it matters

The release is a small, concrete example of what the sovereign-stack argument looks like in practice. A model derived from an openly licensed base, post-trained on curated data, released with open weights, a playground and an API is something other Indian teams can pick up, inspect and build on without asking permission. Publishing the post-training method alongside the weights is the part that makes the result reproducible rather than merely usable, and it is the reason the company framed the announcement as the opening of a series rather than a one-off launch.