Skip to content
Concept

open-weights

open-weights — Concept. definition: Model whose weights are downloadable and runnable (≠ open training pipeline); opens up the self-host option and reversibility, without being equivalent to open-source in the strict sense · geopolitical reading (Ball): Intrinsically decelerationist: deter AI capex and create a "veil of ungovernability" over AI; their generalization would lead to "AI communism"

Downloadable, runnable weights: that is the definition, and its boundary matters. Open weights are not open source AI in the OSI sense, which additionally requires the training code and documentation of the data. Mozilla's The state of open source AI (v1.0.1, July 2026) records the usage shift: a majority of tokens routed on OpenRouter went to open weights by mid-2026, the seven highest-volume models all open, though closed providers still lead by number of requests.

Moonshot AI's Kimi K3, launched 16 July 2026 with a vendor-stated ~2.8 trillion parameters and a one-million-token context, placed frontier capability inside that category. SFEIR's reading treats the benchmark question as secondary. A proprietary model is consumed; an open-weights model is recovered as an option: run it, port it, stop being captive. Kimi K3 joins GLM-5.2 on that ground. The same argument carries the analysis of the Mistral and Microsoft agreement announced 21 July 2026: reserved Azure capacity in Europe and air-gapped execution still rest on an American hyperscaler's infrastructure, and the open-weights character of Mistral's weights is named as the only element that makes that sovereignty genuinely portable.

Dean W. Ball reads the same fact in reverse. Opening weights deters AI capex, which makes the practice decelerationist, and he attributes to self-described accelerationists an appetite for the resulting veil of ungovernability. A world of open weights ends, in his account, in state-run AI as public infrastructure, which he calls dystopian.

Type
Concept
definition
Model whose weights are downloadable and runnable (≠ open training pipeline); opens up the self-host option and reversibility, without being equivalent to open-source in the strict sense
geopolitical reading (Ball)
Intrinsically decelerationist: deter AI capex and create a "veil of ungovernability" over AI; their generalization would lead to "AI communism"
relations
8
Cited in
3 fiches

Neighborhood

Kimi K3 Kimi GLM-5.2

← is an instance of

Kimi K3 TECHNOLOGIE high confidence stable Source ↗
Kimi TECHNOLOGIE high confidence stable
GLM-5.2 TECHNOLOGIE high confidence stable Source ↗

Cited in (3)