Sarvam open-sources 30B, 105B reasoning models

This is less a random Reddit benchmark thread than a meaningful new open-model release, and the real story is Sarvam trying to prove India can ship a full-stack sovereign AI stack with competitive reasoning, coding, and multilingual performance.

• –Sarvam 105B is positioned against models like Qwen3-Next-80B, GPT-OSS-120B, GLM-4.5-Air, and DeepSeek R1 rather than a direct apples-to-apples Qwen 3.5 matchup
• –Sarvam 30B looks especially notable for practical deployment, with the company claiming strong coding and agentic results plus higher inference throughput than Qwen baselines on some hardware
• –The differentiator is not just raw benchmarks but Indian-language performance, tokenizer efficiency across many Indic scripts, and deployment optimization from H100s down to laptops
• –Open weights on Hugging Face and AI Kosh make this more useful to developers than a closed API-only release, especially for teams experimenting with local inference, vLLM, or SGLang
• –Benchmark claims are strong but still company-reported, so the next important signal will be independent comparisons against Qwen 3.5, DeepSeek, and other popular open reasoning models