Reasoning on one graphics card: QwQ-32B and what it unlocks
Alibaba open-sourced a 32-billion-parameter reasoning model this week, reported to match far larger ones. The size is the news.
On 5 March 2025 Alibaba’s Qwen team released QwQ-32B under an open licence: a 32-billion-parameter reasoning model reported to match the performance of models many times its size, including on mathematics and coding benchmarks.
Claims of parity are always worth discounting. The parameter count is not, because it decides where the model can physically run.
Why 32 billion is the interesting number
A model of this size fits, with the usual compression, on a single high-end graphics card. That is a machine under a desk, not a rack in a data centre and not a rented cluster.
Three months ago, running a model that reasons before answering meant paying a provider by the request. It now has a hardware answer, and that answer costs roughly what a good workstation costs.
Who this actually matters to in Switzerland
Not most SMEs. A hosted service under a proper contract remains simpler, cheaper in total, and someone else’s problem at three in the morning.
It matters to the specific cases where the constraint is absolute rather than a preference. A law firm under professional secrecy. A medical practice. An engineering office whose client contract forbids third-party processing outright. A fiduciary handling several clients’ books who does not want to explain a transfer basis to each of them.
For those, the answer to « can we use AI on this material at all? » changed from no to yes-with-a-server sometime in the last six months, and this release moves it further.
The costs that do not appear in the announcement
Hardware is the cheap part. Somebody has to update it, monitor it, secure it and be available when it stops. A model on a machine in your office is a server you now operate, with everything that implies.
And an open model is not automatically good at your work. Test it on your own documents, in your own German, before anyone buys a card.
The reasonable position in March 2025
If you have no hard confidentiality constraint, use a hosted service and revisit in a year. If you do have one, this is worth two days of testing rather than a dismissal, which is more than could have been said last autumn.