SMALL NEEDS A FORKLIFT

The real news

The story, straight

An AI model answers by processing text in chunks called tokens. Parameters are the learned settings behind its answers, not a headcount of tiny employees.

Thinking Machines Lab introduced Inkling 2 Small on Nov 5: 594 billion total parameters, with 17 billion active per token. Its earlier Inkling-Small had 276 billion total and 12 billion active.

The new Small is more than twice the old Small's total size.

Lil'meme is trying to fit this Small into a carry-on. The suitcase has requested structural engineering.

How does something that huge avoid doing all the work every time? A mixture of experts splits work among specialist sections: each token goes to 8 of 384 experts.

That cuts calculation; it doesn't erase the memory needed to hold the whole model. Think hotel: fewer rooms awake, same enormous building.

Artificial Analysis's Intelligence Index gives it 31, as cited in the announcement. That's a test score, not proof it wins every task.

It launched with closed weights through Tinker. Weights are the learned settings needed to run the model yourself.

The Nov 5 announcement says the company intends to release them under Apache 2.0 the following week, after external safety testing and red-teaming. An intention is not a completed release.

The useful lesson: compare active work, memory needs and actual availability, not the adjective on the box. Small is a name, not a luggage allowance.

Like or repost if your carry-on also needs a loading dock.

Thanks for supporting Lil'meme. Your support is appreciated.