Qwen3-VL 30B-A3B
Pandan vs stock vLLM
Time to first token−17.9%
End-to-end latency−8.0%
Throughput / H100+21.8%
Relative GPU cost−17.9%
Realtime inferenceruns on Pandan
We build inference infrastructure for always-on AI models
with low, predictable latency for continuous voice and video streams.
Less waiting. More seeing.
Faster responses and higher throughput than stock vLLM on the same H100, with no drop in measured quality.
Gemma 4 31B
Pandan vs stock vLLM
Time to first token−21.8%
End-to-end latency−30.7%
Throughput / H100+36.2%
Relative GPU cost−26.6%
Same H100 and 20 images per request, with caches off. Latency is p50; throughput is measured at 32 concurrent requests. Stock runs default BF16 serving; Pandan runs tuned FP8. GPU cost is normalized from measured throughput.
Realtime should remember.
Stateless APIs repeatedly upload and process the same frames. Pandan is building around a persistent stream instead.
01 / INGEST
Send each frame once
Reuse recent visual context without paying the network cost again.
02 / MEMORY
Process only what changed
Reuse cached visual context so follow-up questions get faster as the stream and audience grow.
03 / RESPONSE
Respond in the moment
Coordinate vision and voice so responses arrive while the moment still matters.