Erik van Zijst
prutser@mas.to
<p>Victorian Gentleman Programmer</p><p>PGP: <a href="https://keybase.io/erikvanzijst" target="_blank" rel="nofollow noopener" translate="no"><span class="invisible">https://</span><span class="">keybase.io/erikvanzijst</span><span class="invisible"></span></a><br /><a href="http://medium.com/@erikvanzijst" target="_blank" rel="nofollow noopener" translate="no"><span class="invisible">http://</span><span class="">medium.com/@erikvanzijst</span><span class="invisible"></span></a></p>
Posts
-
View post
I have run out of Claude tokens early again this week. Instead, I&#39;m going to try and use the local rig running qwen3.8-27b for what I would otherwise have handed to Opus. Qwen3.8 is slow on Strix Halo, but it is capable. It can&#39;t touch Opus, but it does feel like the first local model genuinely capable of coding in a real codebase.
-
View post
Rerunning the experiment with Qwen3.8-Flash-Next 125b on the same 128GB Strix Halo machine yields an even better result: a Rust port with full feature parity in only 4 hours and only 3 compactions. 262k q8_0 context, Vulkan, low reasoning, preserve thinking. https://mas.to/@prutser/117223965735698905
-
View post
Omarchy has a bad rep on this platform, but daily driving Linux on the desktop does benefit from an AI agent on standby
-
View post
Hahah I noticed Qwen3.8 cursing to itself, but only hidden in its training, not in its output 😂