Qwen 3.7 and the Gospel of Open Weights

There’s a particular kind of excitement that lives in corners of the internet where people argue about quantisation formats and token generation speeds. It is extremely nerdy. It is also, if you care about who gets to use powerful AI tools, genuinely important.

Qwen 3.7 dropped recently, and the announcement sent certain communities into a state that I can only describe as “physiologically enthusiastic.” People were excited about 122 billion parameters, 17 billion active per inference run, something called MTP that apparently doubles generation speed without quality loss, and a 512k context window. Someone tried to explain all of this patiently to a confused commenter, and I appreciated the effort. The explainer was good.

Here’s the bit that interests me, though. Underneath all the hardware flex and benchmark celebration, there’s a real tension sitting quietly in the middle of it.

Alibaba built Qwen. Alibaba is not a charity. For a while, releasing capable open-weight models made strategic sense: crash the market, disrupt the incumbents, build goodwill in the open-source community, and make the established players look expensive. It worked. Qwen models punched well above their weight, and the local LLM community embraced them.

But Alibaba’s priorities appear to be shifting. A few people in the thread noted it plainly: the company has moved from disruption mode toward monetisation. The Max series, the genuinely frontier-level stuff, may not be open-weighted this time around. One commenter put it bluntly: releasing highly capable local models hurts their own revenue. That’s just true. It would be odd if a large commercial enterprise didn’t eventually notice that.

There’s also a counterweight, which is that the Chinese government’s current economic plan apparently includes pushing toward open AI ecosystems. So you have central economic planning, which I’m not going to pretend is great, accidentally serving the interests of people who just want to run a large model on their home hardware. The world is full of these little ironies. I’ve stopped finding them surprising and started finding them faintly amusing.

Someone in the thread mentioned waiting to buy a 512GB Mac Studio specifically to run the 397B model, which may or may not be released as open weights. That’s a significant purchase to make on a rumour. I understand the impulse completely. There’s something about running a genuinely capable model locally, without sending data to someone else’s server, that feels worth chasing. Privacy, control, the satisfaction of it. The Mac Studio idea isn’t crazy if you have the money.

I don’t have that money, and I’m not sure I’d spend it that way if I did. But I get why people do.

The broader point is that the open-weight AI ecosystem is not a stable equilibrium. It exists because a few large players, mostly Chinese labs, found it strategically useful to subsidise the open-source community. If that calculus changes, and the signals suggest it’s changing, then the community that grew up around free, capable, locally-runnable models is going to face some adjustment. Smaller models will still come out. The 9B and 27B class stuff will probably stay open because it doesn’t threaten the commercial tier. But the frontier models? I wouldn’t count on them.

That’s not a catastrophe. It’s just the predictable shape of how this plays out when the early disruption phase ends and the boring business phase begins. The community will adapt. It always does. Someone will train something new, or fine-tune something that exists, or figure out a quantisation trick that makes the previous generation more useful than anyone expected.

I don’t know how this resolves. Nobody does. But I’d rather understand the tension clearly than pretend the open-weight golden era is a permanent feature of the landscape.