Sarvam AIproprietary
Released: 2025-03-01

Saaras v3 (Speech-to-Text)

Frontier speech recognition model supporting code-mixed Hinglish, colloquial dialects, and accented Indian English across 22 languages with high word-error-rate (WER) robustness.

Intelligence Index
82.0/ 100
Calibrated multi-domain composite
Throughput Speed
180tok/sec
Output streaming throughput
Time To First Token
200ms
Initial chunk server latency
LiveBench Contam-Free
71.5%
September 2026 suite score

Sourced Benchmark Evaluations (61)

Verified performance across authoritative benchmarks with provenance tracking.
BenchmarkStatusScoreTrustDateSource TypeProvenance
ARC-AGI-3Active50.7%
93
2025-04-05independentSource ↗
ARC-AGI-2Nearing Saturation60.8%
75
2025-03-29independentSource ↗
FrontierMathActive18.2%
98
2025-03-26independentSource ↗
PutnamBenchActive33.8%
90
2025-03-25independentSource ↗
SWE-LancerDeprecated40.8%
20
2025-03-23independentSource ↗
Omni-MATHDeprecated48.1%
10
2025-03-23independentSource ↗
SoohakActive49.4%
98
2025-03-21independentSource ↗
FrontierCodeActive34.8%
98
2025-03-21independentSource ↗
Humanity's Last ExamActive34.3%
93
2025-03-19independentSource ↗
StructFlowBenchNearing Saturation64.1%
75
2025-03-19independentSource ↗
AIMENearing Saturation61.8%
75
2025-03-19independentSource ↗
OlympiadBenchActive53.3%
80
2025-03-18independentSource ↗
MathArenaActive52%
99
2025-03-17independentSource ↗
FireBenchActive63.4%
95
2025-03-17independentSource ↗
Terminal-BenchActive42%
80
2025-03-17independentSource ↗
OpenAI-MRCRActive68.9%
80
2025-03-16independentSource ↗
SimpleQAActive42.6%
90
2025-03-16independentSource ↗
CodeContestsDeprecated39%
10
2025-03-16independentSource ↗
ComplexBenchNearing Saturation61.9%
80
2025-03-15independentSource ↗
SWE-bench ProActive37.2%
90
2025-03-15independentSource ↗
LongBench-ProActive67.2%
80
2025-03-15independentSource ↗
ZeroSCROLLSNearing Saturation65.6%
78
2025-03-14independentSource ↗
MultiPL-EDeprecated55.2%
10
2025-03-14independentSource ↗
LiveCodeBenchSaturated49.2%
39
2025-03-13independentSource ↗
AgentIFNearing Saturation51.3%
80
2025-03-13independentSource ↗
MATHSaturated63.7%
25
2025-03-13independentSource ↗
InfiniteBenchNearing Saturation69.7%
65
2025-03-13independentSource ↗
MercuryDeprecated52.8%
10
2025-03-12independentSource ↗
SCROLLSNearing Saturation65.6%
65
2025-03-12independentSource ↗
MMLU-ProNearing Saturation75.4%
75
2025-03-12independentSource ↗
RULERNearing Saturation75.4%
84
2025-03-11independentSource ↗
TabMWPSaturated59.8%
25
2025-03-11independentSource ↗
CyberSecEvalDeprecated54%
10
2025-03-11independentSource ↗
SWE-bench VerifiedSaturated46.8%
25
2025-03-10independentSource ↗
MMLU-ReduxSaturated80.4%
25
2025-03-10independentSource ↗
QuALITYNearing Saturation72.2%
78
2025-03-10independentSource ↗
LongBenchNearing Saturation70.5%
65
2025-03-09independentSource ↗
SWE-benchDeprecated40.8%
10
2025-03-09independentSource ↗
MGSMNearing Saturation61.8%
65
2025-03-09independentSource ↗
AGIEvalSaturated67.1%
25
2025-03-08independentSource ↗
QASPERNearing Saturation68.1%
65
2025-03-08independentSource ↗
TruthfulQASaturated72.2%
25
2025-03-08independentSource ↗
GSM-HardDeprecated62.4%
10
2025-03-07independentSource ↗
SuperGLUESaturated79.6%
25
2025-03-07independentSource ↗
CodeXGLUEDeprecated56.4%
23
2025-03-06independentSource ↗
MT-BenchNearing Saturation7.8%
75
2025-03-06independentSource ↗
BIG-Bench HardSaturated76.4%
25
2025-03-06independentSource ↗
WinoGrandeSaturated81.1%
28
2025-03-05independentSource ↗
MBPP+Nearing Saturation61.2%
65
2025-03-05independentSource ↗
IFEvalDeprecated73.9%
25
2025-03-05independentSource ↗
GPQANearing Saturation72.2%
78
2025-03-05independentSource ↗
Needle-in-a-HaystackSaturated96.4%
44
2025-03-04independentSource ↗
HumanEval+Saturated64.8%
25
2025-03-04independentSource ↗
CommonsenseQASaturated81.9%
28
2025-03-04independentSource ↗
MBPPSaturated63.6%
25
2025-03-03independentSource ↗
GSM8KSaturated70.2%
25
2025-03-03independentSource ↗
GLUESaturated84.2%
28
2025-03-03independentSource ↗
MMLUSaturated83.6%
25
2025-03-02independentSource ↗
HumanEvalSaturated68.4%
25
2025-03-02independentSource ↗
ARC-AGISaturated73.3%
25
2025-03-02independentSource ↗
LiveBench (Continuous)Active71.5%
94
2025-02-01independentSource ↗
OpenAI Compatible API ExecutionAPI Ready
from openai import OpenAI

client = OpenAI()
response = client.chat.completions.create(
    model="saaras-v3",
    messages=[{"role": "user", "content": "Evaluate multi-step logic problem"}],
    temperature=0.2
)
print(response.choices[0].message.content)

Evaluation & Sourcing Notes

Scores listed for Saaras v3 (Speech-to-Text) represent verified evaluations extracted from official research papers, independent evaluation suites (HELM, LMSYS, OpenCompass, LiveBench), and verified audit reports.

All benchmarks marked as saturated or deprecated reflect historical performance where the benchmark no longer provides active discriminative power.