Sarvam 105B, the first competitive Indian open source LLM
sarvam.ai
March 7, 2026
30 min read
🔥🔥🔥🔥🔥
56/100
Summary
Sarvam 30B and Sarvam 105B are open-source reasoning models trained from scratch on large-scale, high-quality datasets. The training was conducted in India under the IndiaAI mission, optimizing various aspects including tokenization, model architecture, and execution kernels.
Key Takeaways
- Sarvam AI released two open-source models, Sarvam 30B and Sarvam 105B, trained on large-scale, high-quality datasets entirely in India.
- Sarvam 105B excels in reasoning, programming, and agentic tasks, while Sarvam 30B is optimized for real-time conversational applications.
- Both models achieve state-of-the-art results on Indian language benchmarks, outperforming larger models.
- The models utilize a Mixture-of-Experts Transformer architecture that supports efficient training and deployment across various hardware platforms.
Community Sentiment
Positives
- The initial results of Sarvam 105B are promising for a first model release, indicating potential for future improvements and applications in the AI landscape.
- There is hope for more models in the 30B parameter range, which could enhance the competitive landscape of open-source LLMs.
Concerns
- Sarvam 105B was quickly outperformed by Qwen models, raising concerns about its competitive viability in the current market.
- The model struggles with critical inputs, particularly in moderation tasks, which highlights significant limitations in its practical applications.
- The choice to focus on MXFP4 for Apple Silicon is criticized as misguided, given the lack of hardware support, suggesting poor strategic planning.
