Sarvam AIproprietary
Released: 2025-03-01

Bulbul v3 (Text-to-Speech)

Human-grade expressive conversational speech synthesis engine supporting regional accents, emotional inflections, and natural prosody in 15+ Indian languages.

Intelligence Index
80.5/ 100
Calibrated multi-domain composite
Throughput Speed
150tok/sec
Output streaming throughput
Time To First Token
220ms
Initial chunk server latency
LiveBench Contam-Free
70.0%
September 2026 suite score

Sourced Benchmark Evaluations (51)

Verified performance across authoritative benchmarks with provenance tracking.
BenchmarkStatusScoreTrustDateSource TypeProvenance
ARC-AGI-3Active49.4%
93
2025-04-05independentSource ↗
ARC-AGI-2Nearing Saturation59.3%
75
2025-03-29independentSource ↗
FrontierMathActive16.8%
98
2025-03-26independentSource ↗
PutnamBenchActive31.2%
90
2025-03-25independentSource ↗
Omni-MATHDeprecated44.4%
10
2025-03-23independentSource ↗
SWE-LancerDeprecated37.4%
20
2025-03-23independentSource ↗
SoohakActive45.6%
98
2025-03-21independentSource ↗
FrontierCodeActive31.9%
98
2025-03-21independentSource ↗
AIMENearing Saturation57%
75
2025-03-19independentSource ↗
Humanity's Last ExamActive33.4%
93
2025-03-19independentSource ↗
StructFlowBenchNearing Saturation63%
75
2025-03-19independentSource ↗
OlympiadBenchActive49.2%
80
2025-03-18independentSource ↗
MathArenaActive48%
99
2025-03-17independentSource ↗
FireBenchActive62.2%
95
2025-03-17independentSource ↗
Terminal-BenchActive38.5%
80
2025-03-17independentSource ↗
CodeContestsDeprecated35.8%
10
2025-03-16independentSource ↗
SimpleQAActive41.9%
90
2025-03-16independentSource ↗
SWE-bench ProActive34.1%
90
2025-03-15independentSource ↗
ComplexBenchNearing Saturation60.7%
80
2025-03-15independentSource ↗
MultiPL-EDeprecated50.6%
10
2025-03-14independentSource ↗
AgentIFNearing Saturation50.4%
80
2025-03-13independentSource ↗
MATHSaturated58.8%
25
2025-03-13independentSource ↗
LiveCodeBenchSaturated45.1%
39
2025-03-13independentSource ↗
MercuryDeprecated48.4%
10
2025-03-12independentSource ↗
MMLU-ProNearing Saturation74.1%
75
2025-03-12independentSource ↗
CyberSecEvalDeprecated49.5%
10
2025-03-11independentSource ↗
TabMWPSaturated55.2%
25
2025-03-11independentSource ↗
SWE-bench VerifiedSaturated42.9%
25
2025-03-10independentSource ↗
MMLU-ReduxSaturated78.9%
25
2025-03-10independentSource ↗
SWE-benchDeprecated37.4%
10
2025-03-09independentSource ↗
MGSMNearing Saturation57%
65
2025-03-09independentSource ↗
TruthfulQASaturated70.8%
25
2025-03-08independentSource ↗
AGIEvalSaturated65.4%
25
2025-03-08independentSource ↗
SuperGLUESaturated77.5%
25
2025-03-07independentSource ↗
GSM-HardDeprecated57.6%
10
2025-03-07independentSource ↗
MT-BenchNearing Saturation7.6%
75
2025-03-06independentSource ↗
CodeXGLUEDeprecated51.7%
23
2025-03-06independentSource ↗
BIG-Bench HardSaturated74.5%
25
2025-03-06independentSource ↗
MBPP+Nearing Saturation56.1%
65
2025-03-05independentSource ↗
IFEvalDeprecated72.6%
25
2025-03-05independentSource ↗
WinoGrandeSaturated79%
28
2025-03-05independentSource ↗
GPQANearing Saturation70.8%
78
2025-03-05independentSource ↗
HumanEval+Saturated59.4%
25
2025-03-04independentSource ↗
CommonsenseQASaturated79.8%
28
2025-03-04independentSource ↗
MBPPSaturated58.3%
25
2025-03-03independentSource ↗
GSM8KSaturated64.8%
25
2025-03-03independentSource ↗
GLUESaturated82.1%
28
2025-03-03independentSource ↗
MMLUSaturated82.1%
25
2025-03-02independentSource ↗
ARC-AGISaturated71.4%
25
2025-03-02independentSource ↗
HumanEvalSaturated62.7%
25
2025-03-02independentSource ↗
LiveBench (Continuous)Active70%
94
2025-02-01independentSource ↗
OpenAI Compatible API ExecutionAPI Ready
from openai import OpenAI

client = OpenAI()
response = client.chat.completions.create(
    model="bulbul-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 Bulbul v3 (Text-to-Speech) 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.