Benchmark Performance

Faster inference with no loss in model quality — see how a model served through the BLACKBOX AI API matches its official benchmark scores.

Serving a model through the BLACKBOX AI API gives you higher throughput and lower latency without sacrificing model quality. Our inference optimizations (custom kernels, tuned batching, and GPU-level scheduling) speed up how fast tokens are generated — they do not change the weights or the behavior of the model itself, so accuracy on the benchmarks the model already performs on stays intact.

Terminal Bench 2.1 — GLM 5.2 Max

The chart above compares GLM 5.2 Max on Terminal Bench 2.1. Running on BLACKBOX AI scores 78.1% accuracy, matching — and slightly edging out — the 77.9% of the official reference implementation. In other words, you get the full speed benefits of our inference stack with zero decrease in benchmark accuracy.

Throughput vs other providers

Speed on BLACKBOX AI is significantly higher than other inference providers — and, as shown above, that speed comes with no loss in model performance.

Measured in tokens per second, BLACKBOX AI leads at 424.6 t/s, ahead of Databricks (356.4 t/s), Fireworks (330.2 t/s), Baseten (323.8 t/s), and Wafer (224.1 t/s). You get the fastest generation available without giving up any of the benchmark accuracy shown above.