The Open Weights Question

The argument about open weights is conducted in the language of principle. It rests on an empirical claim — about how far behind open systems are — that is rarely measured.

Both sides of this debate make a quantitative assumption. Those who favour open release generally assume the capability gap is large enough that open systems pose no marginal risk. Those who oppose it generally assume the gap is small enough that open release effectively transfers frontier capability. Both cannot be right, and neither is usually measured.

Measuring the gap in time

We measure the gap as elapsed time rather than as a score difference, because score differences are not comparable across instruments and because time has a natural interpretation: how long after a capability appears in a closed system does it appear in an openly released one.

The gap narrowed through 2024 and has widened modestly since. Our reading is that the widening reflects the growing share of frontier capability that derives from post-training investment — reinforcement learning at scale, extended reasoning curricula, and evaluation infrastructure — which is harder to replicate from released weights than pre-training is.

The gap is task-dependent

Aggregating to a single number conceals an important structure. On single-pass tasks — classification, extraction, translation, short-form generation — the gap is small and has at times been negligible. On tasks requiring extended reasoning, tool use over many steps, or reliable long-horizon coherence, the gap is roughly twice as large.

Licensing has fragmented

A separate finding, and one with practical consequences: the word 'open' has substantially detached from its licensing meaning in this domain. Fewer than half of the widely used releases described as open in 2026 carry a license that meets the Open Source Definition. Restrictions on field of use, on downstream deployment scale, and on competitive application are common.

We take no position on whether open release is desirable. We observe that the empirical premises on both sides of the argument are testable, that they have testable answers, and that the answers are more nuanced than either position typically allows: the gap is real, persistent, and much larger for some capabilities than others.