Latest news with #Sarvam-M.

Indian Express

Sarvam AI debuts flagship open-source LLM with 24 billion parameters

Indian AI startup Sarvam has unveiled its flagship Large Language Model (LLM), Sarvam-M. The LLM is a 24-billion-parameter open-weights hybrid language model built on top of Mistral Small. Sarvam-M has reportedly achieved new standards in mathematics, programming tasks, and even Indian language understanding. According to the company, the model has been designed for a broad range of applications. Conversational AI, machine translation, and educational tools are some of the notable use cases of Sarvam-M. The open-source model is capable of performing reasoning tasks like math and programming. According to the official blog post, the model has been enhanced through a three-step process – Supervised Fine-Tuning (SFT), Reinforcement Learning with Verifiable Rewards (RLVR), and Inference Optimisations. When it comes to SFT, the team at Sarvam curated a wide set of prompts focused on quality and difficulty. They generated completions using permissible models, filtered them through custom scoring, and adjusted outputs to reduce bias and cultural relevance. The SFT process trained Sarvam-M to function in both 'think', which is complex reasoning, and 'non-think' or general conversation modes. On the other hand, with RLVR, Sarvam-M was further trained using a curriculum consisting of instruction following, programming datasets, and math. The team used techniques like custom reward engineering and prompt sampling strategies to enhance the model's performance across tasks. For inference optimisation, the model underwent post-training quantisation for FP8 precision, achieving negligible loss in accuracy. Techniques like lookahead decoding were implemented to boost throughput; however, challenges in supporting higher concurrency were noted. Notably, in combined tasks with Indian languages and math, such as the romanised Indian language GSM-8K benchmark, the model achieved an impressive +86% improvement. In most benchmarks, Sarvam-M outperformed Llama-4 Scout, and it is comparable to larger models like Llama-3.3 70B and Gemma 3 27B. However, it shows a slight drop (~1%) in English knowledge benchmarks like MMLU. The Sarvam-M model is currently accessible via Sarvam's API and can be downloaded from Hugging Face for experimentation and integration.