Thinking Machines Lab Releases Inkling-Small with Improved Training

Thinking Machines Lab has released the full weights of Inkling-Small, a new model available for fine-tuning on Tinker and for various chat functionalities on Tinker Playground. Inkling-Small's training process incorporated improvements over its larger counterpart, including changes to its pre-training data mix and machine learning recipe. Post-training with on-policy distillation helped it surpass Inkling on reasoning and agentic coding benchmarks.