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

13 pointsposted 6 hours ago
by har-ki
(har-ki.github.io)

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

6 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

31 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

5 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