Thinking Machines Lab Releases Inkling — Their First Open-Weights Multimodal Model
Thinking Machines Lab has released Inkling, its first open-weights model — a heavy multimodal foundation model with 975B total parameters (41B active), 1M token context, trained on 45T tokens across text, images, audio, and video. A smaller preview, Inkling-Small (12B active), is available alongside.
The company is open that Inkling doesn’t top every leaderboard. Instead, the focus is on providing a capable multimodal base that can be fine-tuned for specific tasks — and the highlight is Tinker, a platform for customizing Inkling directly. In a demo, the model wrote its own fine-tuning job, learned to avoid the letter “e” in responses, ran an eval, and switched to new weights autonomously. It’s a full customization loop: model → task → data → fine-tune → eval → new weights. Thinking Machines appears to be building a factory for bespoke models tailored to specific products.