Coding models locally is easier when the hardware can hold them. The RTX 3090 Ti has 24 GB of VRAM and the Coder quantization cuts the model size enough to leave room for runtime overhead. Strata measures throughput directly on your machine, so the reported 2255 reads per second and 93 writes per second reflect real usage.
Hardware fit and VRAM limits
The 24 GB of VRAM is the decisive factor for which quantizations run without swapping. Coder was chosen because it keeps the model compact while preserving enough context for coding tasks. With Strata loading the model, VRAM usage stays inside the 24 GB boundary.
How Strata handles local inference
A local inference benchmark shows the speed of the Strata engine on a single RTX 3090 Ti. Reading throughput reaches 2255 tokens per second, writing throughput 93 tokens per second, while the card draws about 600 W under load. Strata reports these figures in real time so you can verify performance yourself.
| Metric | Value |
|---|---|
| VRAM used | 24 GB |
| Read TPS | 2255.1 |
| Write TPS | 93.1 |
| Power draw | 600 W |
Core term defined
"Coding quantization" means a low-precision format that keeps the bulk of a language model's weights in int8 while keeping a small number of parameters in higher precision for better reasoning in code contexts.