RTX 5090 vs RTX 4090

On the matched pairs cited here (same model, same quant, llama.cpp, chat-short), the RTX 5090 decoded 1.31× to 1.73× as fast as the RTX 4090: Qwen3-Coder-30B-A3B Q4_K_M at 304.9 vs 232.5 tok/s (≈1.31×, r_vsype4pm1ko vs r_1k8942tw3zf); LFM2.5-24B-A2B Q4_K_M at 308.5 vs 231.9 (≈1.33×, r_c8k6fleym09 vs r_bqmpv9kncyd); gpt-oss-20b Q4_K_M at 324.2 vs 242.6 (≈1.34×, r_dwp5fgdapmw vs r_t-a9v_ipdit); Qwen3.6-35B-A3B Q4_K_M at 224.0 vs 160.6 (≈1.39×, r_g1aw3mizd32 vs r_m6j1zxwvp-f); Qwen2.5-Coder-14B Q4_K_M at 137.3 vs 89.87 (≈1.53×, r_xnr_k4emoyr vs r_eqnsqmgz29a); Qwen3.6-27B Q4_K_M at 75.09 vs 47.14 (≈1.59×, r_pob-co51bd- vs r_d_wda_rw62f); Qwen3-32B Q4_K_M at 69.83 vs 43.47 (≈1.61×, r_vhbqyt1f_xr vs r_ayy9rjgha5r); and Qwen3.8-27B Q4_K_M at 4K context at 75.19 vs 43.38 (≈1.73×, r_qfrb2zmyk0k vs r_7cap38a021-; different GGUF publisher, lmstudio-community vs ggml-org). Ratios are calculated from single measured runs. llm-speed measured the 5090 runs other than Qwen3.6-35B-A3B, and the 4090 runs for LFM2.5, Qwen3.6-35B-A3B and Qwen3.8-27B; the other runs are community-submitted, and each run page says which. One of the 5090 quants (Qwen3.6-35B-A3B) comes from the model file name; the gpt-oss-20b, Qwen3-Coder-30B-A3B, Qwen2.5-Coder-14B, Qwen3-32B and Qwen3.6-27B pairs use the labelled Q4_K_M core-set runs. Only the LFM2.5 pair used the same llama.cpp build; the others ran on different machines and builds, so these are measured reference points, not a controlled ranking. Nemotron Bolt still sits in the unmatched table (quant unreported on both cards), and so does gemma-4-12b-it-qat Q4_0: the 5090 run used a 16K context and the 4090 run 8K (r_v5arqmazf31 vs r_jfh0qkngw76), so they are not paired. With MTP on, a community-submitted 4090 run measured 84.51 tok/s on Qwen3.8-27B (r_7j8rpk9mwhg); there is no 5090 MTP run yet, so don't compare MTP-on against MTP-off across cards.