In an interview published October 25, 2026, INT21 said its agent swarm built a working Rust inference runtime for Fish Audio’s Drama 3 about an hour after the model became publicly accessible, without early access. An inference runtime is software that executes a trained model; INT21 described work on the existing Drama 3 model, not training a new one.
How INT21 says the runtime was developed
INT21 describes a loop in which agents generate candidate implementations, test them against a baseline on target hardware, keep changes that work, then repeat the process. Engineers set the constraints, target hardware and acceptance criteria.
The company says its Rust runtime has no Python interpreter in the serving path. INT21 also identified Drama 3’s FP8 mixture-of-experts, codebook multi-token prediction and linear attention among the model mechanisms involved in the work.
What INT21’s three-prompt test measured
INT21 reports lower time to first audio and real-time-factor values, and higher throughput, for its Rust runtime on each of three prompts compared with SGLang-Omni. Time to first audio measures the wait before the first speech chunk. Throughput is the number of audio frames generated per second. The real-time factor (RTF) compares processing time with the duration of the generated audio; INT21 treats lower values as better.
Each runtime configuration used two NVIDIA B200 GPUs, one for generation and one for the codec. The test used one speech stream, a 1-frame first chunk and 20 warm trials per prompt.
| Prompt | Runtime | Time to first audio (ms) | Throughput (audio frames/s) | Real-time factor |
| Hello | INT21 Rust | 15.11 | 223.4 | 0.0964 |
| Hello | SGLang-Omni | 37.79 | 136.9 | 0.1573 |
| Multiple sentences | INT21 Rust | 17.55 | 249.2 | 0.0864 |
| Multiple sentences | SGLang-Omni | 38.61 | 154.9 | 0.1390 |
| Voice reference | INT21 Rust | 23.58 | 231.0 | 0.0932 |
| Voice reference | SGLang-Omni | 38.03 | 145.4 | 0.1481 |
What the benchmark says about Python
The comparison is between two complete runtime implementations. It does not isolate removing Python as the cause of the measured differences. The reported results cover one speech stream, three prompts and the two-B200 setup described above.
A separate hands-on test by Tonbi’s AI Garage ran unquantized Drama 3 locally through SGLang Omni on an NVIDIA DGX Spark, demonstrating voice cloning and direction tags. That test used SGLang Omni, not INT21’s Rust runtime.
