Qwen Team (2024) — ARXIV-2412.15115
Key takeaways
- The Qwen2.5 family, open-weight across a wide size range, which is what makes it the default base for method papers that need a model they can fine-tune, perturb or run evolution strategies over
- Filed for what it is, not for what it recommends: when a result here says “Qwen2.5-7B”, this is the document that says which model that was
Standing in the anthology
Carries no practice, deliberately — ADR-032. A model report is
admissible for what it establishes; this one establishes an identity, and the
identity is the thing forty-five documents in this record have been assuming.
The incumbent shape, for the third time this week. The record already held
LIT-182 (Qwen3), LIT-135 (Qwen3.8) and LIT-125 (Qwen2.5-Coder) — the
successors and a sibling — and not the base model they and everyone else
measure against. Same pattern as DoReMi behind the mixing laws, and DCLM
behind the filtering practices: the thing that becomes everybody’s baseline
gets cited constantly and filed never, because nobody encounters it as news.
Unread — no NOTE. ADR-025 makes that mean exactly what it says, and for a
model report the identity is what the record needed.