The small Qwen that was promised for the same week has no repository, licence or date
Qwen3.8-27B was committed for the week of 10 August. It has no official repository, no model card, no licence file and no benchmarks of its own, and third-party trackers now describe it as delayed with no new date given.
The 2.4T flagship is the headline and the 27B is the one that mattered to most people. A 27-billion-parameter model runs on hardware a developer already owns; a 2.4-trillion-parameter sparse model runs in a datacentre. The release that shipped is the one fewer people can use.
Absence of a repository is not the same as absence of a model, and the sequencing may be nothing more than staged rollout. But the specific pattern — flagship ships, accessible sibling slips, no revised date — is worth tracking because it has a shape that repeats. Announced-then-delayed small models have been a recurring feature of open-weight releases this year, and each instance is individually explicable.
What makes it non-trivial is what the small model does competitively. Frontier-scale open weights change what a research lab can inspect; a good 27B changes what a startup can afford to run. Only one of those pressures a hosted API business.
Recorded as pending, not as cancelled. The distinction matters, and so does noting that nobody has committed to a new date.
OrcaRouter — Qwen3.8-27B Open Weights: What We Know Before the Drop → · LLM Stats — AI Updates Today (August 2026) →