Mira Murati's lab unveiled its first open-weights model. The interesting part is the bet: a base you can download and fine-tune for yourself.
- Thinking Machines released Inkling on 15 July 2026 - the lab's first model, with open weights on Hugging Face.
- 975 billion parameters total, 41 billion active; context up to 1 million tokens; trained on text, image, audio, and video.
- The company itself writes that this isn't the strongest model today - they're pitching it as a base for fine-tuning through their platform, Tinker.
Thinking Machines is the lab of Mira Murati - the former chief technical officer of OpenAI. For a long time it showed off ideas and tools, but not a model of its own. Now it has one. It's called Inkling and it comes with open weights - meaning anyone can download it and change it, not just rent it through someone else's cloud.
Don't fall for the headline. Half the model news out there is fighting for one line on a leaderboard, and here the company itself says it isn't leading the leaderboard - and that's on purpose. The bet is different: not to rent you the smartest brain, but to give you a base you take home and bend to your own work. For someone who builds things, that's more interesting than any benchmark. Open weights plus a fine-tuning tool means the model can become YOURS, instead of staying someone else's black box that gets swapped out from under you tomorrow.
Here's what I'm taking away for us. I'll leave the numbers in the table to marketing - they change every two months. What matters is whether an open-weights base you can fine-tune for a specific task does a better job than the general rental model. That gets proven in practice - on our data, our task. Until then, Inkling is a promising base, not a finished answer. We'll try it, then we'll tell you.