The new RTX 4090 24GB vs RTX 5090 32GB local-AI comparison separates the capacity decision from the generation decision. RTX 5090 adds eight gigabytes of nominal device memory and moves from Ada capability 8.9 to Blackwell capability 12.0, but it also raises NVIDIA reference board power from 450W to 575W.
Where 24GB stops and 32GB continues
Qwen’s official Qwen2.5-32B-Instruct-GGUF repository makes the boundary concrete. Q4_K_M is 19.9GB, leaving 4.1GB of file-only margin on RTX 4090 and 12.1GB on RTX 5090. Q5_K_M is 23.3GB, leaving an unsafe 0.7GB on the 24GB card but 8.7GB on the 32GB card. Q6_K is 26.9GB: it exceeds RTX 4090 nominal capacity and leaves 5.1GB on RTX 5090. Q8_0 is 34.8GB and fits wholly on neither.
These are file-size boundaries, not peak-memory measurements. KV cache, runtime workspace, modality, display use, concurrency and allocator behavior remain outside those margins. Use the 32B VRAM guide or enter the exact workload in the LLM VRAM calculator.
Newer and larger are not complete benchmark results
NVIDIA publishes 1,008GB/s memory bandwidth for RTX 4090 and 1,792GB/s for RTX 5090, along with 16,384 versus 21,760 CUDA cores. Those specifications do not establish prompt-processing rate, generation rate, time to first token, energy per completed request or value. A matched comparison must hold model revision, quantization, runtime/version, kernel path, context, cache, modality, prompt/output profile, batch, concurrency and timing constant.
If RTX 4090 spills weights or cache to system memory while RTX 5090 remains resident, report the different placement rather than presenting the result as a pure architecture comparison. Record exact board, power limit, clocks, display use, exposed memory, peak device memory, CPU offload, system RAM and wall power.
Continue the decision
Read the full comparison · Compare three generations · Run and submit a matched test · Stage a private watchlist
Published October 25, 2026. Sources include NVIDIA’s official RTX 4090 and RTX 5090 specifications and Qwen’s official Qwen2.5-32B-Instruct-GGUF repository. HighVRAM has not independently benchmarked this pair.