Qwen 3.8 27B — MTP or DFlash?

<!-- 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>With practically all of LocalLlama glazing Qwen 3.5/3.6 for it's coding skills. Along with the fact that Alibaba themselves are focusing on

<!-- SC_OFF --><div class="md"><p>Title.</p> <p>I have the draft for Qwen3.5 (not 3.6) 27B, would it be compatible? I tried this combination in oMLX and PP speed is actually much

<!-- SC_OFF --><div class="md"><p>Really loving Qwen 27b more than any other llm from when I can remember. It works so well. Having 48gb vram can anyone recommend any other

<!-- SC_OFF --><div class="md"><p>Does anyone know if MTP works with more than one 3090' yet? I see the 5090's talking about it, but would like to know for us poors.</p>

<!-- 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>

<a href="https://news.ycombinator.com/item?id=49953495">Comments</a>

<!-- SC_OFF --><div class="md"><p>I'd jump on runpod and ssh in to test my workloads, but they don't have it.</p> <p>Would love to know how well this runs, particularly as context