RTX 5090 dominates local AI benchmarks

The RTX 5090’s GDDR7 bandwidth finally makes sparse MoE models feel like local "instant" intelligence, but Apple’s memory architecture still wins on utility for deep codebase reasoning.

• –The 5090 delivers a ~30% generation speed increase over the 4090, peaking at 240 t/s during long-context generation.
• –Qwen3.6-35B-A3B activates only 3B parameters per token, allowing the aging RTX 3090 to still deliver a respectable 140 t/s.
• –Mac M5 Max is restricted by memory bandwidth for raw speed (~95 t/s) but can natively host 1M token context windows that would require 4+ RTX 3090s to fit in VRAM.
• –These results suggest that for developer agentic workflows, the 5090 is the new gold standard for latency, while high-RAM Macs remain the standard for large-scale repo analysis.