Faster inference without an observed loss in model quality. That is the promise of the BLACKBOX AI API — and it is a claim you can check.
Serving a model through our API gives you higher throughput and lower latency with no observed loss in model quality. Our inference optimizations — custom kernels, tuned batching, and GPU-level scheduling — speed up token generation without retraining or changing the model weights. Benchmark results can still vary with evaluation-harness configuration, sampling, and run-to-run variance. These comparisons support preserved quality; score differences should not be attributed to serving optimizations.
Why serving speed need not cost accuracy
Most speedups in inference come from how tokens are computed and scheduled, not what is computed. We accelerate the serving path with three levers:
Custom kernels — hand-tuned GPU code that executes model operations with far less overhead.
Tuned batching — grouping and scheduling requests to keep the GPUs saturated while serving the same model and request settings.
GPU-level scheduling — placing work across the hardware so tokens stream out faster end to end.
None of these retrain or alter the model weights. To evaluate the result, we ran public benchmarks against our deployments and compared them with the published reference scores. Because harness settings and run-to-run variance can shift results, we use these comparisons to check for quality regressions rather than claim that the inference stack makes a model more accurate.
GLM 5.2
On Terminal Bench 2.1, our GLM 5.2 deployment scored 78.1% accuracy, compared with the 77.9% official reference. This close result supports no observed quality regression; the 0.2-percentage-point delta may reflect differences in evaluation setup or run-to-run variance and should not be attributed to serving.
Running on BLACKBOX AI vs official reference. Higher is better.
| Model | % |
|---|---|
| BLACKBOX AI | 78.1% |
| Official reference | 77.9% |
Terminal Bench 2.1. Observed scores can vary with evaluation setup and run-to-run variance.
NVIDIA Nemotron Ultra
On Terminal Bench, our NVIDIA Ultra deployment scored 55% against the published NVIDIA Ultra result of 53.9%. The 1.1-percentage-point difference may reflect harness configuration or run-to-run variance and should not be attributed to serving optimizations.
Our deployment vs the official NVIDIA Ultra score. Higher is better.
| Model | % |
|---|---|
| BLACKBOX AI | 55% |
| NVIDIA Ultra (official) | 53.9% |
Terminal Bench. Observed scores can vary with evaluation setup and run-to-run variance.
Kimi K2.7 Code
Under our Terminal Bench evaluation, the Kimi K2.7 deployment scored 71.9% against the published 67% reference. Differences in harness configuration, sampling, and run conditions can produce score gaps. This comparison provides no evidence of a quality regression, but it should not be interpreted as an accuracy gain from serving.
Our deployment vs the official Kimi K2.7 score. Higher is better.
| Model | % |
|---|---|
| BLACKBOX AI | 71.9% |
| Kimi K2.7 (official) | 67% |
Terminal Bench. Observed scores can vary with evaluation setup, sampling, and run conditions.
The takeaway
Across three benchmark comparisons — GLM 5.2, NVIDIA Ultra, and Kimi K2.7 — the BLACKBOX AI deployments produced scores at or above the published references under our evaluation conditions while delivering higher tokens-per-second and lower latency. These results show no observed accuracy regression; they do not establish that serving optimizations improve accuracy.
The reason is simple: we optimize the serving path, not the model. Custom kernels, tuned batching, and GPU-level scheduling make each model faster to run while leaving its weights unchanged.