What's this about?
Multiverse Computing released Quasar 438B, billed as Europe's leading AI model. The top HN comment congratulates them on the milestone while noting the bar: it scores below Qwen3.8-27B — a model with roughly one-sixteenth the parameters.
Why would a 438B model lose to a 27B one?
Because parameter count stopped being the determining variable some time ago. Data quality, training compute actually spent, the training recipe, and post-training all matter more, and Qwen's recent models are widely regarded as unusually well-executed. A large parameter count with a mediocre training run produces an expensive model that is not very good. The comparison is not a knock on the ambition; it is a measurement of how far ahead the frontier is.
What's the community saying?
HN 162 points, 102 comments
- The politest possible framing of the result: congratulations on the milestone, hopefully next iterations get better.
- A serious disagreement about whether provenance matters: "as a European, I don't care where an open-weight model comes from" versus "I do, because I'm not oblivious to geopolitics or digital sovereignty."
- A long, admiring digression about Qwen3.8-Flash-Next fitting into 128GB at Q4 and getting close to state of the art — which is the actual comparison the thread cares about.