the RTX 5080 is newer, faster on paper, and the wrong pick for local AI. here's the 8GB reason why

The short answer

For local AI, the RTX 4090's 24GB beats the RTX 5080's 16GB despite the 5080 being newer: 24GB runs 32B-class models at Q4 that simply don't fit in 16GB. The 5080 is faster on models both can run…

16GB vs 24GB is the decision. The 5080 tops out around 14B-class models; the 4090 runs 32B-class (e.g. Qwen3-32B at Q4, ~19–20GB). No amount of Blackwell speed loads a model that doesn't fit.

On shared workloads the 5080 is competitive-to-faster: ~94 t/s vs the 4090's ~70 t/s on Qwen3-14B Q4 at 16k (hardware-corner, llama.cpp).

The 5080 wins on FP4 support, power efficiency, and image/video generation (compute-bound work where Blackwell pulls ahead).

Street prices are close enough (~$1,400 vs ~$2,200 tracked) that the 4090's premium buys real capability, not just nostalgia — if you can still find one.

Buy the 5080 for ≤14B models plus gaming/image-gen; buy the 4090 for the 32B-class local models that are the sweet spot of 2026.