Post #1561167

Also, with regards to running models locally, wow, what a rabbit hole I have bene in. To summarize it best: for the Strix Halo Qwen3 Coder Next is the best model so far when it comes to consistency and token generation. From my short experience, anything >= 20 t/s is good, so as long as quality is good as well (which this model yields!). But for planning it's still better to use a SOTA model and then have the smaller model execute.