🆕Qwen 4 is coming, and we’re ready for it
A lot of people are searching for Qwen 4 right now. Alibaba has said it’s in training, and the talk is that it could drop any day. When it does, I want it running inside Claude Code on a regular computer the same day, with no cloud and no API key. That’s what Claude Code Local is for.
As of today the weights are not out. So this is the plan, written down before the release, so you can hold me to it.
📥Day one: download it
The minute Qwen 4 shows up on the official Qwen page on Hugging Face, I pull it down, plus the MLX builds for Macs and the GGUF builds for Linux and Windows as the community posts them. I’ll take every size that fits on the machines people actually own, from a 16 GB MacBook Air up to a 128 GB Mac.
🧰Make it work inside Claude Code
Claude Code expects to talk to Claude. Claude Code Local runs a small server on your own machine that speaks the same language and translates each model’s way of calling tools into what Claude Code understands. It already does this for Qwen 3.5, Qwen 3.8 and Gemma 4. If Qwen 4 changes how it formats tool calls, that’s the first thing I fix.
📏Test it fairly, then post the numbers
Every model here goes through Agent-12, the same coding test the current pick took. Qwen 3.8 27B scored 12 out of 12 on the easy set and 7 out of 8 on the hard set. Qwen 4 gets the same test at the same size class, run the way Alibaba says to run it. Then I post memory used, tokens per second, and which Mac or PC each size needs.
If it wins, it becomes the default. If it doesn’t, I’ll say so. Nothing on the page will claim Qwen 4 works until I’ve run it myself.
⭐How to catch it
Star or watch Claude Code Local on GitHub. The day Qwen 4 lands, the README gets the results and a one-line install. If you hit your Claude Code usage limit before then, it already keeps you going on Qwen 3.8 or Gemma 4 today: type exit, then keep going.

