March 7, 2026
Code meets culture war
Sarvam 105B, the first competitive Indian open source LLM
India drops a mega AI — and the comments go nuclear
TLDR: Sarvam released a large Indian-made open-source AI model powering its Indus assistant, and it’s already live to try. The comments erupted over alleged political bias in the system prompt, whether it’s truly open, and if more local models mean real innovation—or just louder meme wars.
Sarvam just open-sourced two big chatbot brains, including the 105B model powering its Indus assistant, and the internet did what it does best: argue loudly. Fans cheered an Indian-built, open-source model trained entirely in-country and already live via API and on Hugging Face. But the vibe quickly shifted from “Congrats, India!” to “Wait, what did it just say?”
One lightning rod: a screenshot of the system prompt telling the model not to adopt certain charged terms from foreign media. Cue a flame war over bias and censorship inside AI. Supporters framed it as legal caution; critics called it nationalist guardrails. Meanwhile, a snarky nationality jab sparked pushback and mod warnings, while others tried to refocus on whether “open source” here is truly open and how it stacks up to the big US labs.
Amid the drama, pragmatists asked the only question that matters: Does it work? Early testers traded jokes (“ask it your doubts?”) and benchmarks, while one thoughtful thread argued that more homegrown models mean a richer idea space, not just copycat tech. Verdict from the crowd: huge milestone, spicy politics, and meme-fueled chaos—aka a normal day on the internet.
Key Points
- •Sarvam AI open-sourced two LLMs, Sarvam 30B and Sarvam 105B, trained from scratch on in-house curated data across pre-training, SFT, and RL.
- •Training was conducted entirely in India using compute provided by the IndiaAI mission.
- •Both models use a Mixture-of-Experts Transformer with sparse expert routing and long-context features; 30B uses GQA, while 105B adds depth and MLA.
- •The models are in production: 30B powers the Samvaad conversational platform; 105B powers the Indus assistant for complex reasoning and agentic workflows.
- •Weights are available via AI Kosh and Hugging Face, with API access and local inference support using Transformers, vLLM, and SGLang.