Thinking Machines Lab sets out staged release framework for open weights

Thinking Machines Lab published “A safe path to open weights”, setting out a staged framework for deciding when a model’s weights should be released in full. The company proposed four stages of increasing access ahead of any public weight release: inference-API access, so defenders can probe the model and prepare countermeasures; fine-tuning access through its Tinker platform, for vetted researchers to build safety layers on top of it; white-box access for safety researchers to study the model’s internals and failure modes; and monitored public access to observe how the model is actually misused before deciding on full release. The company stressed that the stages were not a fixed pipeline that automatically ends in open weights — progression depends on evidence gathered at each stage, and a model can be held back indefinitely.

The framework was explicitly tied to the company’s own releases. Thinking Machines said it had applied this reasoning to Inkling and Inkling-Small, its two open-weight models, concluding that neither extended the frontier of dangerous capability — in areas such as cybersecurity or biological and chemical weapons design — beyond what existing open-weight models from other labs already posed, and that full release was therefore justified.

The post was notable as a rare instance of a lab publishing a general decision procedure for open-weight release, rather than a one-off risk assessment for a single model, at a time when the debate over open-weight AI had increasingly turned on whether any release process could keep pace with capability. It set an explicit standard — no incremental risk beyond the existing open-weight frontier — that other labs’ own claims about open releases would now be measured against.