INT21's AI swarm built a Rust runtime for Drama 3 in an hour, beating SGLang-Omni

Startup INT21 used its SwarmOS AI agent swarm to build a custom Rust inference engine for Fish Audio's Drama 3 speech model in about an hour, with no early access. On two Nvidia B200 GPUs, time to first audio was 15.11–23.58 ms versus 37.79–38.61 ms for SGLang-Omni, with 1.59–1.63x higher throughput.

• Rust runtime built in about 60 minutes without early access to Drama 3
• Time to first audio: 15.11–23.58 ms vs 37.79–38.61 ms for SGLang-Omni
• Throughput 1.59–1.63x higher on 2x Nvidia B200 GPUs
• Swarm has already built over 20 engines across 7 model types