Thinking Machines Readies an Open Model, Carefully

Thinking Machines Lab, the startup run by former OpenAI chief technology officer Mira Murati, released a new model on November 5 and then did something its rivals in open AI rarely do. It kept the weights to itself, for now.

The model, Inkling 2 Small, is available first through the company's Tinker API. The lab says it intends to publish the full weights under the permissive Apache 2.0 licence in the coming weeks, but only after outside safety testing and red-teaming are finished. No firm date is attached to that release.

That ordering is the point. In a companion post titled A Safe Path to Open Weights, the lab argues that a model can be shipped behind an interface, studied, and only then handed over in full. It is a middle road between the two camps that have defined the debate: labs that keep their best models closed, and labs that drop the raw weights the moment a model is ready.

What is under the hood

Inkling 2 Small is a sparse mixture-of-experts model with 594 billion total parameters and 17 billion active at any one time, according to coverage from MarkTechPost. It takes both text and images, and it lets developers dial the amount of reasoning effort up or down depending on the task.

On the Artificial Analysis Intelligence Index, a composite benchmark, it scores 31. That puts it ahead of other recent open-weight releases such as Mistral's Large 4 at 26 and Meta's Muse Spark Open at 25, though it still trails the strongest closed frontier systems.

A quieter kind of release

The staggered approach stands apart from the pattern set this autumn, when groups like Moonshot AI and Mistral published their weights on day one. Thinking Machines is betting that a short gap for evaluation costs little and buys some insurance against misuse.

Whether the weights actually arrive, and when, now depends on how the testing goes. The lab has committed to the licence but not the calendar, so the open release is a promise rather than a shipment. For a company that has spent its first year defining what careful looks like, that may be the message as much as the model.