Sarvam AI
Open-source 105B-parameter MoE reasoning chat model with an OpenAI-compatible API, long-context support, coding, agentic, and Indian-language strengths.
Sarvam-105B is a 105B MoE open-weights reasoning model with a documented OpenAI-compatible API, 128k context, and stated Indian-language specialization. Its $0.042/M input and $0.17/M output estimate is materially more economical than most frontier hosted models and supports a cost score above Grok 3 (70) and near DeepSeek-V3.2 (91). However, Artificial Analysis’ Intelligence Index of 12 and the absence of independent broad capability results keep technical capability below DeepSeek-V3.2 (82), Kimi K2.5 (83), and frontier anchors. Coding evidence is mixed. The official model card reports 45.0 on SWE-Bench Verified using SWE-Agent, but Artificial Analysis reports only 1.5% on TerminalBench Hard; this supports a coding/agentic score below Hermes 4.3 36B (55) and far below GLM-4.6 (79) or DeepSeek-V3.2 (84). A 0.710 Arena-Hard-v2 score at rank 10 provides some preference-based support for general reasoning, but it is not a substitute for independently replicated knowledge or reasoning benchmarks. No confirmed LiveCodeBench leaderboard entry was retrieved. Developer usability benefits from open weights, Hugging Face distribution, and OpenAI-compatible access, placing it slightly above GLM-4.6 on interface maturity. Multimodal capability is not evidenced in supplied materials. Only Sarvam is tracked as an API provider, and no measured throughput, latency, or response-time data is available, limiting availability and speed confidence. Official benchmark reporting is useful but should not outweigh the sparse independent evidence.
