Sarvam AI opens access to Sarvam-M LLM via API and Hugging Face download

Sarvam AI's 24-billion-parameter Sarvam-M, rivaling Meta and Google models, is now publicly available for developers and researchers. With significant improvements in Indian languages, math, and coding, this open-weight model invites broad experimentation and integration.

Sources:
NewsBytesIndianexpress
Updated 2h ago
Tab background
Sources: NewsBytesIndianexpress
Bengaluru-based startup Sarvam AI has officially launched Sarvam-M, a 24-billion-parameter multilingual Large Language Model (LLM) designed to excel in Indian languages, mathematics, and programming tasks.

Built on the open-weight Mistral Small model by French firm Mistral AI, Sarvam-M is a hybrid-reasoning, text-only model that supports two operational modes: a "think" mode for complex logical reasoning, math, and coding, and a "non-think" mode for general-purpose conversation.

The model has demonstrated remarkable performance improvements, including a 20% average boost on Indian language benchmarks, a 21.6% enhancement in math-related tasks, and a 17.6% increase in coding benchmarks. Notably, on combined tasks such as the romanised Indian language GSM-8K benchmark, Sarvam-M achieved an 86% improvement.

Sarvam AI claims that Sarvam-M outperforms Meta's LLaMA-4 Scout on most benchmarks and rivals larger models like LLaMA-3.3 70B and AI21 Labs' Gemma 3 27B.

The development process involved a three-step enhancement: Supervised Fine-Tuning (SFT), Reinforcement Learning with Verifiable Rewards (RLVR), and Inference Optimisations, ensuring both accuracy and efficiency.

Sarvam-M is now accessible to developers and researchers via Sarvam's API and is available for download on Hugging Face, facilitating experimentation and integration into various applications.

"Sarvam-M is a versatile model supporting both complex reasoning and efficient general conversation," the company noted.

This launch marks a significant milestone in India's AI landscape, offering a powerful, open-weight multilingual model tailored to regional languages and advanced computational tasks.

Sources: NewsBytesIndianexpress
Bengaluru-based Sarvam AI has launched Sarvam-M, a 24-billion-parameter multilingual Large Language Model (LLM) built on Mistral Small. The model, accessible via API and Hugging Face, shows significant improvements in Indian languages, math, and coding benchmarks, outperforming several leading AI models.
Section 1 background
Sarvam-M is a single, versatile model that supports both 'think' and 'non-think' modes. The think mode is for complex logical reasoning, mathematical problems, and coding tasks, while the non-think mode is for efficient general-purpose conversation.
Sarvam AI Official Blog
NewsBytes
Key Facts
  • Sarvam AI, a Bengaluru-based startup, has launched Sarvam-M, a 24-billion-parameter multilingual hybrid-reasoning Large Language Model based on the open-weight Mistral Small model from French firm Mistral AI.NewsBytesIndianexpress
  • Sarvam-M was enhanced through a three-step process: Supervised Fine-Tuning (SFT), Reinforcement Learning with Verifiable Rewards (RLVR), and Inference Optimisations to improve performance.Indianexpress
  • Sarvam-M set new benchmarks with a 20% average improvement on Indian language tasks, 21.6% improvement on math tasks, and 17.6% improvement on coding benchmarks compared to the base model.NewsBytesIndianexpress
  • On combined Indian language and math tasks such as the romanised GSM-8K benchmark, Sarvam-M achieved an 86% improvement.Indianexpress
  • Sarvam AI claims that Sarvam-M outperforms Meta's LLaMA-4 Scout on most benchmarks and rivals larger models like LLaMA-3.3 70B and Google's Gemma 3 27B.NewsBytes
  • Sarvam-M is now publicly accessible via Sarvam's API and is available for download on Hugging Face for experimentation and integration.Indianexpress
Key Stats at a Glance
Parameter size of Sarvam-M LLM
24 billion parameters
NewsBytes
Average improvement on Indian language benchmarks
20%
NewsBytes
Improvement on math-related tasks
21.6%
NewsBytes
Improvement on coding benchmarks
17.6%
NewsBytes
Improvement on combined Indian language and math tasks (GSM-8K benchmark)
86%
Indianexpress
Article not found
⌂ Home ‹

Source Citations