Benchmark MiniMax-M2.5 on 8*H20 perf test

<!-- SC_OFF --><div class="md"><p>Every output token the model generates requires reading the entire KV cache from High Bandwidth Memory.</p> <p>Here's what that actually looks

<!-- SC_OFF --><div class="md"><p>Wrote a deep dive on <strong>FlashAttention-4 (03/05/2026)</strong> that's relevant for anyone thinking about inference performance.</p>

<table> <tr><td> <a href="https://www.reddit.com/r/LocalLLaMA/comments/1t4uj9h/mtp_on_strix_halo_with_llamacpp_pr_22673/"> <img alt="MTP on strix halo with llama.cpp (PR #22673)"

<!-- SC_OFF --><div class="md"><p>Running Llama-3-70B with long contexts and hitting HBM limits constantly. 128K context burns approx 40GB just for KV cache — half the H100 gone

<!-- SC_OFF --><div class="md"><p>I could not find good data points on what speed one could get with a single 5090 and enough DDR4 RAM. </p> <p>My system: AMD EPYC 7532 32core

<!-- SC_OFF --><div class="md"><p>So i recently updated LMstudio after a long pause and updated my llama.cpp runtimes too.. i was shocked.. i thought maybe something like

<!-- SC_OFF --><div class="md"><p>I'm genuinely unsure if I'm doing something wrong, or if the model is just worse, despite the community enthusiasm. I tend to use it with pi

<table> <tr><td> <a href="https://www.reddit.com/r/singularity/comments/1s7rcwr/qwen35_omni_qwens_latest_generation_of_fully/"> <img alt="Qwen3.5 Omni - Qwen’s latest generation