A Safe Path To Open Weights

Thinking Machines proposes widening access to powerful models in stages, using each stage to gather evidence and prepare defenses before publishing downloadable weights. Once weights are public, the release cannot be undone, and refusal training cannot be relied on as a safeguard because users can customize a model to remove it.

Hosted access offers intermediate options. Researchers can test models, and developers can customize them through a managed training service, while the provider retains monitoring and access controls. The proposal makes further opening conditional on evidence about both model capabilities and defensive readiness. It leaves the precise thresholds for advancing or stopping to a future, more detailed framework.