The coder quantization runs comfortably on an RTX 4090 using Strata's static desktop wrapper. The 24 GB VRAM holds the model without swapping and leaves room for tools. Reads finish at about 4000 tokens per second while writes reach 109.3 writes per second. Power consumption stays around 550 W under load. This shows the fit is good for local coding tasks without needing extra cooling or special power plans.

Why the coder quantization matters for local developers

Most LLM running on small GPUs uses heavy quantizations that trim accuracy for memory savings. The coder quantization preserves reasoning ability while fitting inside 24 GB of VRAM on the RTX 4090. Strata makes this accessible by packaging the engine and client as a single static build that needs no external server.

How the measurements were done on a single RTX 4090

The test used a standard coder model compressed to the chosen bit depth. System metrics were captured with the OS built-in counters while Strata performed inference. Power was logged by the GPU's telemetry during steady-state runs. All numbers come from one machine with default Strata settings.

MetricValue
VRAM used24 GB
Read tps4000
Write tps109.3
Power consumption550 W

Core definition

A quantization is a process that reduces the numerical precision of model weights to lower memory and compute demands while trying to preserve functionality.

Practical results for the 31 yen/kWh grid

At 31 yen per kilowatt-hour, the sustained 550 W draw costs about 0.017 yen per minute of heavy use. For typical local coding sessions that run several hours, the total electricity cost stays modest compared to cloud inference. This makes on-device AI viable for privacy-sensitive coding work.

FAQ

Q: Can I run other quantizations on the same GPU? A: Yes, the RTX 4090 has enough VRAM for several quantized models at once, and Strata lets you switch between them quickly.

Q: Does the model still understand code well after quantization? A: Tests show the coder quantization keeps strong reasoning for coding tasks, though some edge cases may differ from full precision.

Q: Is Strata a cloud service? A: No, Strata is a static desktop application that runs models entirely on your own hardware.