NVIDIA RTX 5090 Redefines Local AI Performance, Outpaces MacBook Pro and API Costs

A new build featuring an NVIDIA GeForce RTX 5090, equipped with 32GB of VRAM and 64GB of RAM running Ubuntu, has demonstrated exceptional performance in local AI model execution, establishing a new benchmark for on-device inferencing. Utilizing LM Studio, tests on models like Gema 4 (7.5B parameters) and Quen 3.6 (35B parameters) yielded token generation speeds of approximately 190 tokens per second (TPS) and 253 TPS, respectively. The system efficiently handled Quen 3.6, consuming 20-22GB of VRAM while supporting a maximum context window of over 260,000 tokens, highlighting the critical role of dedicated VRAM for larger models.

The RTX 5090 system’s capabilities extend beyond raw speed, demonstrating practical application in software development workflows. Integrating Quen 3.6 via an LM Studio API into a custom web application for model comparison revealed that the local AI generated a clicker game in approximately half the time (1 minute) compared to an API call to GPT 5.5 (2 minutes), with zero cost. Further integration with code agents like Pi and Visual Studio Code (via a dedicated LM Studio extension) showed the local model efficiently resolving code issues and automating file modifications, such as adding a new URL redirect in seconds. The multimodal capabilities were also highlighted, allowing the local AI to process images for tasks like web page development from screenshots or generating descriptive alt text. In a direct comparison, a MacBook Pro M4 Max (36GB unified memory) running an MLX-optimized Quen 3.6 achieved around 76-78 TPS, underscoring the RTX 5090’s significant performance lead, often by a factor of three or more.

Cost analysis positioned the RTX 5090 PC at €5,199, a desktop configuration that offers substantial dedicated VRAM, versus a comparable MacBook Pro M5 Max (hypothetically €4,849) or a Mac Studio M4 Max (€3,000, but without M5 Max yet). While the Apple machines offer portability and integrated design, the NVIDIA setup delivers superior raw AI performance due to its dedicated memory architecture. This makes the RTX 5090 a compelling investment for professionals and small businesses heavily engaged in local AI development, model training, and privacy-sensitive operations, where the return on investment can be rapidly realized through reduced API costs and accelerated development cycles, as evidenced by a custom model training task decreasing from one hour on a Mac to under ten minutes on the new PC.