Supporting case
Ling's base weights broaden adaptation access, but still require major compute
What happened
inclusionAI's October 27 Ling-3.1 releases include a base checkpoint for continued training and domain adaptation and a separate Flash-1019 model. The base uses an MIT licence. It has 560 billion total parameters, with about 25 billion active per token; activating only part of the network reduces computation but does not make all the stored weights disappear. A base checkpoint is not itself a ready-to-use assistant.
The Flash-1019 deployment instructions specify configurations such as eight H200 or four GB300 GPUs. That illustrates who can realistically use the access: well-resourced operators, rather than ordinary small organisations on a laptop. Published benchmark claims remain provider evaluations. Deployment documentation ↗.
Permissive adaptation rights widen model access, while major memory and GPU requirements limit who can exercise those rights. Open weights do not establish small-institution operating capacity.
Dates and implementation stage
Development: 2026-10-27
