RTX 5090 vs RTX 4090

On five matched pairs (same model, same quant, llama.cpp, chat-short), the RTX 5090 decoded faster every time, by 1.33× to 1.75×: LFM2.5-24B-A2B Q4_K_M at 308.5 vs 231.9 tok/s (≈1.33×, r_c8k6fleym09 vs r_bqmpv9kncyd); Qwen3.6-35B-A3B Q4_K_M at 224.0 vs 160.6 (≈1.39×, r_g1aw3mizd32 vs r_m6j1zxwvp-f); gemma-4-12b-it-qat Q4_0 at 141.1 vs 97.72 (≈1.44×, r_v5arqmazf31 vs r_jfh0qkngw76); Qwen3.6-27B Q4_K_M at 73.96 vs 47.14 (≈1.57×, r_f7ulllqu4vk vs r_d_wda_rw62f); and Qwen3.8-27B Q4_K_M at 75.19 vs 43.06 (≈1.75×, r_qfrb2zmyk0k vs r_11kn1epxewh). Ratios are calculated from single measured runs. Two of the 5090 quants come from the model file name. Only the LFM2.5 pair used the same llama.cpp build; the others ran on different hosts and builds, so these are measured reference points, not a controlled ranking. Five more models with runs on both cards (gpt-oss-20b, Qwen3-Coder-30B-A3B, Qwen2.5-Coder-14B, Qwen3-32B and Nemotron Bolt) are listed separately, because the 5090 runs don't record a quant. With MTP on, the 4090 measured 84.51 tok/s on Qwen3.8-27B (r_7j8rpk9mwhg); we have no 5090 MTP run yet, so don't compare MTP-on against MTP-off across cards.