M2 Ultra/Qwen3.8 Flash Next Update - latest oMLX introduces substantial speedup

<table> <tr><td> <a href="https://www.reddit.com/r/LocalLLaMA/comments/1w0c0o1/m2_ultra_reporting_in_so_far_so_good/"> <img alt="M2 Ultra Reporting in - so far so good."

<!-- SC_OFF --><div class="md"><p>Running oQ4e+MTP on oMLX 0.7.0, with still more to optimize.</p> <p>Prefill is 1,878 toks.</p> <p>I saw some other benchmarks below what id

<!-- SC_OFF --><div class="md"><p>Right from the oven with the <a href="https://github.com/jundot/omlx/commit/58b3ca549ab7aba075ecd5f1481911e01d819702">latest commit</a>:</p>

<table> <tr><td> <a href="https://www.reddit.com/r/LocalLLaMA/comments/1w2nz07/qwen_38_flash_next_locally_on_simple_mobile_phone/"> <img alt="Qwen 3.8 Flash Next locally on simple

<!-- SC_OFF --><div class="md"><p>PR #22673 (commit 4f13cb7) landed MTP speculative decoding in mainline llama.cpp on May 16. I tested it on two separate rigs.</p> <p>Qwen3.6 27B,

<!-- 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

<!-- SC_OFF --><div class="md"><p>If you're using MTP with Qwen Flash Next and llama.cpp, you can switch to:</p> <p>quants: <a

<!-- 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