Thinking Machines publishes staged open-weight release safety framework alongside Inkling-Small

Thinking Machines Lab published 'A Safe Path to Open Weights' (July 31, 2026), detailing its safety-testing methodology (internal/external/adversarial fine-tuning assessment) for the Inkling model family and a staged-access framework that widens weight availability only as defenses mature, released alongside the smaller Inkling-Small (276B/12B active) open-weight model.

First named vendor to operationalize a graduated, evidence-based release protocol for open-weight models balancing dual-use misuse risk against openness, offering a template other open-weight labs (Meta, Mistral, Moonshot) may be pressured to adopt amid rising scrutiny of open-weight jailbreak risk.

AI labs and enterprises evaluating open-weight model provenance/safety before fine-tuning or self-hosting; policy teams tracking open-weight governance norms.