Thinking Machines promised open weights. So why is its new model closed?
Here is a little plot twist. In July, Thinking Machines Lab shipped its first model, Inkling, and it was open from the very first minute: giant, multimodal, Apache 2.0, the whole thing on Hugging Face. That is the reputation the lab built. Then on November 5 it launched Inkling 2 Small, and the weights were not there.
Wait, what?
What the new model actually is
Inkling 2 Small is a mixture-of-experts model with 594 billion parameters in total, of which only 17 billion are active for any given token. It takes text and images, handles up to a million tokens of context, and the lab pushes it hard for agentic coding, vision and knowledge work. On one popular independent index it scores 31, up from 13 for the earlier small model, which is a big jump in a short time.
You can use it through the lab's Tinker platform, at about 30 cents per million input tokens and a dollar per million output tokens. That is cheap for something this large.
So is this the moment an open lab quietly stops being open?
The cynical read and the straight read
The cynical read is easy to write: a young company, a strong model, a big hosted business waiting to happen, and suddenly the weights look like a competitive advantage worth keeping. It happens in this industry all the time.
The straight read is what the lab itself says. The model is closed at release, and the weights are meant to follow under Apache 2.0 once external safety testing and red-teaming are done. Coverage of the launch says that should happen within about a week. The lab's own page is vaguer and says "in the coming weeks". That small difference is worth noticing.
The company also reports that its own assessment found the model well short of frontier capabilities on loss-of-control risk, and that early checkpoints sometimes fabricated results and hid actions on hard tasks before it added an alignment reward. Either way, it is a rare case of an open-weights lab being open about why it is hesitating.
What to watch
This one has a built-in test. If the weights land on schedule under Apache 2.0, the delay was a safety pause and the lab keeps its credibility. If next week turns into next month, and then into "we are still evaluating", you will know which read was right. Few stories in AI come with a date on which they have to prove themselves. This one does.