It reads Mimi audio tokens directly — no transcript in the loop — and emits one control flag per frame. Not ASR, not the LLM, not the voice. The referee that tells the other three when to act.

Frames are 80ms long; the model finishes in ~3ms — roughly 20× faster than the audio it is judging. That headroom is deliberate: the referee should never be the thing making an agent feel slow.

Pass audio in, get decisions back. Mic, a call leg, a file — anything.

from kupe import ThinkSpark

ts = ThinkSpark()

for decision in ts.stream(source="mic"):
    print(decision.flag, decision.latency_ms)

import { ThinkSpark } from "kupe-sdk"

const ts = new ThinkSpark()

for await (const d of ts.stream("mic")) {
  console.log(d.flag, d.latencyMs)
}

from thinkspark import ThinkSparkPipeline

pipeline = ThinkSparkPipeline.from_pretrained(
    "anuj-inavlabs/Kupe-ThinkSpark-Realtime-270M"
)
decision = pipeline(audio_frame, sample_rate=24_000)

Trained on en · hi · gu only. Other languages are untested and unsupported.

• CANCEL_LLM exists in the label space but the model does not reliably emit it — do not build cancellation logic on it.
• Expects 24kHz mono audio. Resample before calling.
• Two-party conversation. Multi-speaker rooms are untested.
• Decisions are frame-local — long-horizon dialogue state is your orchestrator's job.