An open reasoning model you could run yourself. November ends with an option
Alibaba published QwQ-32B-Preview under a permissive licence at the end of November 2024. Three months ago this class of model existed only behind an API.
At the end of November 2024, Alibaba’s Qwen team published QwQ-32B-Preview, a 32-billion-parameter reasoning model released under a permissive Apache 2.0 licence. It is aimed at mathematics, logic and coding, and its weights are downloadable.
Three months ago, a model that reasons before answering existed only as a paid service behind somebody else’s API. It is now a file you can host.
What actually changed
Not the leaderboard. The shape of the option.
For most SMEs the honest answer is still: use a hosted service. Running a 32-billion-parameter model needs real hardware, someone to maintain it, and a reason. « Because we can » is not a reason, and self-hosting has a habit of becoming the most expensive way to save money.
But the option now exists where it did not, and that matters in two specific cases.
Genuine confidentiality constraints. A law firm, a medical practice, an engineering office under a client NDA that forbids third-party processing. Until now the answer was « then you cannot use this ». Now there is an answer that involves a server.
Negotiation. Even if you never self-host, a credible alternative changes what a renewal conversation sounds like.
What has not changed
Open weights do not mean free. Hardware, electricity, updates and the person who fixes it at 7am are all real costs, and for a 40-person company they usually exceed an API bill by a wide margin.
Open weights also say nothing about quality on your work. A model tuned for competition mathematics may be unremarkable at reading a Swiss delivery note.
Closing November
This month produced a reported plateau in scaling, a European provider worth evaluating, a connection standard, agents inside Microsoft 365, and two capable reasoning models outside the usual labs.
The pattern across all of it is the same: capability is spreading outward and downward, and the scarce thing is no longer access to a good model. It is knowing which of your own processes is worth pointing one at.