Author: Guy Halfon
The past week has produced the expected chorus of celebration. Moonshot’s Kimi K3 arrived as a 2.8-trillion-parameter mixture-of-experts model with a million-token context window and native multimodality, posting results that sit uncomfortably close to-or in some coding and agentic evaluations ahead of-recent closed systems. Days earlier, Thinking Machines Lab released Inkling, a fully open-weights multimodal model with 975 billion total parameters (41 billion active). The narrative writes itself: open models, and Chinese open models in particular, have closed the gap. The fortress of frontier labs is cracking. Democracy of intelligence has arrived.
That narrative is comforting. It is also largely wrong.
Velocity Is the Point
The error is temporal. People look at the scoreboard on the day of release and treat it as a static ranking. In this field the ranking is already obsolete by the time the press releases land. While the open community trained, evaluated, and shipped Kimi K3 and Inkling, the closed labs did not stand still. They continued training the next generation on infrastructure that most organizations cannot even approximate.
Elon Musk’s old observation that “the factory is the product” applies with unusual force here. The decisive advantage is no longer a single model checkpoint. It is the industrial system that can produce the next checkpoint faster, at larger scale, with tighter feedback loops. xAI’s Colossus clusters-now operating at gigawatt scale and continuing to expand-illustrate the point. The same is true, in different institutional forms, at Anthropic and OpenAI. These organizations have built manufacturing systems for intelligence. The models that emerge from them are downstream of that capacity.
Open releases can and do match last quarter’s frontier. They almost never match the frontier that is already being trained while the celebration is underway. The gap does not close; it moves. And the organizations that own the factories keep moving it.
The Cyber Consequence
This dynamic has immediate implications for security.
Anthropic’s Mythos-class models (Mythos 5 and the related Fable 5) demonstrated capabilities that forced the U.S. government into export controls and tightly limited access. The concern was straightforward: systems this capable at vulnerability discovery, exploit chaining, and autonomous offensive operations lower the cost of sophisticated attacks. Restricting them was an attempt to keep those capabilities inside a trusted perimeter.
The attempt did not work as intended. Within weeks, Chinese laboratories and open-weight efforts produced systems approaching comparable performance on many of the same axes. Domestic chip programs, architectural workarounds, limited and carefully managed access to restricted GPUs, and sheer scale of engineering effort all contributed. The result is predictable. Capabilities that were meant to remain restricted are now available, in open or near-open form, to anyone with the resources and intent to run them.
Defenders who assumed a durable asymmetry-Mythos-level tools for the good guys, something weaker for everyone else-now face a different reality. Attackers will have access to models that are close enough in agentic coding, long-horizon reasoning, and offensive security tasks to matter. The cost of sophisticated reconnaissance, vulnerability research, and payload development continues to fall. The defensive advantage that was supposed to come from privileged access to the most capable systems is eroding in real time.
What This Means for Risk
For those of us who spend our days thinking about third-party and supply-chain risk, the lesson is uncomfortable but clear. Model capability is no longer a reliable differentiator between well-resourced defenders and determined adversaries. The relevant questions shift upstream: who controls the factories, how fast those factories iterate, and whether the organizations we depend on have the operational maturity to absorb continuous, high-velocity capability shocks.
Celebrating the latest open release as a victory against closed models mistakes a snapshot for a trajectory. The trajectory favors the people who own the industrial base. That base is not evenly distributed, and it is not standing still.
