INT21 Says AI Agents Rebuilt Fish Audio’s Drama 3 Runtime in About 60 Minutes

Infrastructure startup INT21 says a swarm of AI agents built a working inference engine for Fish Audio’s Drama 3 model in about 60 minutes. The company used Rust for the customized runtime, according to Wccftech.

INT21 describes its approach as “Self-Improving Infrastructure” and says it uses compute to improve compute. The reported work focuses on software runtimes rather than buying more NVIDIA B200 GPUs. The source frames those efforts as orthogonal, but does not provide a direct benchmark or quantify the runtime’s speed improvement.

The project also highlights INT21’s use of Rust instead of Python for this runtime, with AI agents contributing to its construction. Wccftech characterizes the agents as outperforming human experts, but the available account does not give details of the comparison or identify the experts.

The report says the team produced a viable inference engine in roughly an hour. It does not specify the hardware used, the evaluation method, or how the runtime performs across other models. The reported result therefore concerns this Drama 3 implementation; the source provides no evidence that it generalizes to other workloads.