/series/running-qwen3.6/Recent content in Running Qwen3.6 on ZenithiaHugoen-nzTue, 19 May 2026 08:14:02 +0300/posts/local-qwen-2/Tue, 19 May 2026 08:14:02 +0300/posts/local-qwen-2/<p>It has been almost 10 days since <a href="https://zenithsal.cloudmillgames.com/posts/local-qwen/" target="_blank">Part 1</a> and I&rsquo;ve learned a lot, tweaked and benchmarked more times than I can recall, and used Qwen3.6 in as many contexts as I could in my daily dev life.</p> <h2 id="the-quality">The Quality</h2> <p>I know and trust Qwen3.6-35B-A3B at this point. I&rsquo;ve used it for:</p> <ul> <li>coding features in my current project (Go)</li> <li>implementing and expanding test suites for multiple projects</li> <li>writing build/CI configs for multiple projects to run on my Forgejo server</li> <li>reviewing code as I&rsquo;m writing it</li> <li>implement my own MCP tools to run on my server, also written in Go</li> <li>wrote terminal tools/scripts that I now use daily</li> </ul> <p>I&rsquo;ve been mostly using it through opencode, and sometimes pi.</p>/posts/local-qwen/Fri, 08 May 2026 14:41:45 +0300/posts/local-qwen/<p>My interest in LLMs for coding assist tooling started before the virality of GPT. Back then I tried this extension in vscode called Tabnine which provided &ldquo;smart&rdquo; auto-completion while writing code. It was heavy, vscode felt sluggish and the fans whirred louder, but the thing felt like magic sometimes. Half the time, the suggested completions were completely off, but the other half it felt like it was reading my mind. I enjoyed this experience because there was no friction between accepting or ignoring these completions and it sped up going from idea to implementation.</p>