Benchmark Lab // Hardware fit dashboard

Know what your machine can run before you load it.

Benchmarks turns model picking into evidence. Celestos reads your CPU, RAM, GPU, VRAM, utilization, and temperature, then compares models by score, speed, memory, context, parameters, and license.

See Models Check Requirements

Use cases // Better model decisions

Stop guessing which model belongs where.

Test whether a model fits in your VRAM before loading it

Compare three models on benchmark scores and speed

Pick the fastest model that is good enough for Standard tier

Reserve Elite tier for the model that wins hard reasoning tests

Decide whether a task should stay local or move to cloud

What gets measured

Hardware CPU cores, RAM, GPU, VRAM, utilization, temperature

Speed tokens per second and response behavior

Fit what fits comfortably, what is tight, what should use cloud

Benchmarks MMLU-Pro, HumanEval, MATH, GPQA, AIME, SWE-Bench, Arena Hard

Model metadata VRAM, context, parameters, license, benchmark source

Compare mode // Up to three models

See the trade-offs in one table.

The benchmark view lets you compare models side by side, then use the winner immediately in Chat or route it into a tier. The goal is not to crown one universal best model. It is to find the best model for your hardware and the work in front of you.

Hardware fit test

GPU             RTX 4090
VRAM Capacity   24GB
GPU Util        18%
Active Model    Qwen 14B

Compare         MMLU   HumanEval   VRAM
Qwen 14B        81.1   85.4        10.7GB
Gemma 12B       72.3   78.0         8.0GB
Phi-4 Mini      68.0   71.0         2.5GB
        

How to use the result

Use the strongest model for Elite reasoning.

Use the fastest capable model for Standard tasks.

Use tiny local models for background helpers and edge devices.

Move oversized jobs to Cloud Hub instead of forcing local hardware.