Running Claude Code Offline on an M3 Pro with Qwen3.6

(har-ki.github.io)

13 points | by har-ki 6 hours ago ago

10 comments

  • InTheArena 3 hours ago

    Most reports I ahve seen have 27b (which is dense) being much more performant then 35b (MoE). Is there a reason for 35b?

    • har-ki 2 hours ago

      I am using qwen3.6:35b-a3b-coding-nvfp4, active 3b. Easier on my hardware - Mac M3 Pro.

  • tasuki 5 hours ago

    I understand why local, I don't understand why Claude Code. Is it better than the other harnesses?

    I thought the main reason for using Claude Code was that it was the only harness one was allowed to use with the Claude subscription plan...

    • msradam 5 hours ago

      The inertia of switching. At least for me, being able to maintain the same UX and configuration for agentic coding is a benefit. I have access to internal models at work where I use a LiteLLM proxy with Claude Code, so no Anthropic models, but I can maintain the same .claude between personal and work machines. Though I'm definitely open to a non-Anthropic harness if it's a clear win.

      • papascrubs 11 minutes ago

        Doing anything special to handle the syncing of .claude? Kind of a pain point for me jumping between Windows/Linux and having to deal with path differences.

      • har-ki 4 hours ago

        Yes!! And your LiteLLM-proxy-to-internal-models setup is a cleaner setup than mine.

    • har-ki 4 hours ago

      Fair question. I haven’t benchmarked it against other harnesses, so I can’t say it’s better.

      My reason for using claude code is continuity - I use it for dev (it’s also the tool of choice for most people I work with), and it already has the context. And it supports both frontier and OSS models.

  • corporealshift 5 hours ago

    i don't understand why people would ever want to read this kind of article. The information is ok, but it's a regurgitation of many sources, some of it incorrect, covered in ai slop prose.

    Please share the cool thing you're doing in a blog post, but don't have AI write it for you; it's hard to read and doesn't respect the reader

  • har-ki 6 hours ago

    [flagged]