A model called Small holds 594 billion parameters, and you still cannot download it

Cover image: AI-generated

Hey,

A model called Small holds 594 billion parameters.

That is the name Thinking Machines Lab gave its new model, Inkling 2 Small. It came out on November 5. Small is the second version of the Inkling family, and it is more than twice the size of the first one, which had 276 billion parameters.

Only a small part of it works at a time. For each word it handles, it uses 17 billion parameters, and leaves the rest quiet. That is how it keeps the work moving, even though the full model is huge.

It scored 31 on the Artificial Analysis Intelligence Index. That is up from 13 for the first Inkling Small. The same test puts Mistral Large 4 at 26. If the weights come out as promised, this would be the highest scoring open model built outside China on that test.

Here is the part to watch. You cannot download it yet.

Right now you can only try it on the lab's own platform, Tinker. The full weights are promised under an Apache 2.0 licence next week, after outside safety tests and red team checks finish. No date is fixed. Until then, it is open in plan only.

There is one more catch in the fine print. The score came at a cost. Artificial Analysis says the model used about 150,000 output tokens for each test task, close to four times what the first version used. It is cheap for each token, and dear for each full answer.

The lab built it for long jobs that run for hours, like coding, reading images, and knowledge work. Mira Murati leads the lab. She used to be the chief technology officer at OpenAI.

I like the promise here, and I am holding the praise back until the download link is real. A model that is open next week is not open today.