Sarvam 105B performs strongly on multi-step reasoning benchmarks, reflecting the training emphasis on complex problem solving. On AIME 25, the model achieves 88.3 Pass@1, improving to 96.7 with tool use, indicating effective integration between reasoning and external tools. It scores 78.7 on GPQA Diamond and 85.8 on HMMT, outperforming several comparable models on both. On Beyond AIME (69.1), which requires deeper reasoning chains and harder mathematical decomposition, the model leads or matches the comparison set. Taken together, these results reflect consistent strength in sustained reasoning and difficult problem-solving tasks.
“Unveiling Inefficiencies in LLM-Generated Code.” arXiv, 2025.，更多细节参见新收录的资料
Nature, Published online: 04 March 2026; doi:10.1038/s41586-026-10212-4，更多细节参见新收录的资料
Медсестра занялась сексом с пациентом и обвинила его в изнасиловании02:03
All use our new imatrix data. See some improvements in chat, coding, long context, and tool-calling use-cases.