Sarvam 105B

Why teams choose it

Tradeoffs to know

• Published benchmark results are primarily vendor-reported and should be validated with an independent task-specific eval.
• Thinking mode is on by default in Sarvam docs, so small max_tokens settings can be consumed by reasoning tokens.
• Global model comparison should account for language mix, latency region, context length, and deployment constraints.

When not to use this

• Self-hosting outcomes depend on hardware, quantization, and ops maturity—budget time beyond swapping an API hostname.
• May demand more instrumentation than SaaS-managed APIs to duplicate latency, failover, and support guarantees.
• Benchmark prompts and regressions continuously before rewriting entire routing tables around weights.

Technical specs

Inputs

text

Outputs

text

Capabilities

Indian-language chat, Reasoning, Coding, Long-context document analysis, Tool use, Agentic workflows, OpenAI-compatible chat completions

License

Apache 2.0

Model string

sarvam-105b

Benchmarks

{
  "source": "https://www.sarvam.ai/blogs/sarvam-30b-105b",
  "math500": 98.6,
  "mmlu_pro": 81.7,
  "tau2_avg": 68.3,
  "aime_2025": 88.3,
  "browsecomp": 49.5,
  "gpqa_diamond": 78.7,
  "vendor_reported": true,
  "live_code_bench_v6": 71.7,
  "swe_bench_verified": 45,
  "aime_2025_with_tools": 96.7,
  "indian_language_win_rate_avg": "90%"
}