Faster inference with no 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 without sacrificing model quality. Our inference optimizations — custom kernels, tuned batching, and GPU-level scheduling — speed up token generation, but they do not change the model weights or behavior. The math the model runs is bit-for-bit the same. So benchmark accuracy stays intact, and in practice often edges ahead of the reference numbers.
Why speed doesn't cost you 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 computes the exact same operations with far less overhead.
Tuned batching — grouping and scheduling requests to keep the GPUs saturated without altering any single request's output.
GPU-level scheduling — placing work across the hardware so tokens stream out faster end to end.
None of these touch the model weights. That is the whole point: the model produces the same distribution of outputs it always did, just sooner. To prove it, we ran the same public benchmarks against our deployments and compared them to the official reference scores.
GLM 5.2
On Terminal Bench 2.1, our GLM 5.2 deployment scores 78.1% accuracy — matching, and slightly edging out, the 77.9% of the official reference implementation. Full inference-speed benefits, zero decrease in benchmark accuracy.
Running on BLACKBOX AI vs official reference. Higher is better.
| Model | % |
|---|---|
| BLACKBOX AI | 78.1% |
| Official reference | 77.9% |
Terminal Bench 2.1. Same model weights, accelerated inference. +0.2 pp vs the official reference.
NVIDIA Nemotron Ultra
On Terminal Bench, our NVIDIA Ultra deployment scored 55% against the official NVIDIA Ultra result of 53.9% — again ahead of the reference, with none of the quality trade-off that faster serving is often assumed to require.
Our deployment vs the official NVIDIA Ultra score. Higher is better.
| Model | % |
|---|---|
| BLACKBOX AI | 55% |
| NVIDIA Ultra (official) | 53.9% |
Terminal Bench. Same weights served on the BLACKBOX AI inference engine. +1.1 pp vs the official score.
Kimi K2.7 Code
On Terminal Bench, our Kimi K2.7 deployment scored 71.9%, well above the official Kimi K2.7 result of 67%. The gap comes entirely from serving quality and consistency — the model itself is unchanged.
Our deployment vs the official Kimi K2.7 score. Higher is better.
| Model | % |
|---|---|
| BLACKBOX AI | 71.9% |
| Kimi K2.7 (official) | 67% |
Terminal Bench. Same weights served on the BLACKBOX AI inference engine. +4.9 pp vs the official score.
The takeaway
Across three independent benchmarks — GLM 5.2, NVIDIA Ultra, and Kimi K2.7 — the BLACKBOX AI deployment matches or beats the official reference scores while delivering higher tokens-per-second and lower latency. Faster inference, same intelligence.
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, but the weights — and therefore the answers — stay exactly the same.