Thinking Machines released a new open AI model on Friday. It is called Inkling 2 Mini.
The lab was founded by Mira Murati, a former OpenAI executive. Open means anyone can download the model's weights and run it.
Inkling 2 Mini has 26 billion parameters. A parameter is one small number a model learns during training. More parameters usually mean a bigger, slower model.
The lab's last model, Inkling 2 Small, had 594 billion parameters. Many users said that model was too big to run at home.
Inkling 2 Mini runs on a Mac with 32 gigabytes of memory. It also runs on Nvidia's small DGX Spark computer.
The model loops its middle layers three times per word. Looping lets it think more before it answers. It still writes fewer words to get there.
The announcement tweet earned 4,993 likes in hours. It drew 227 replies and quotes.
The model's page on Hugging Face, a site that hosts AI models, earned 323 likes. Users downloaded it 136 times within a day.
Why did this catch fire? Many builders want strong AI models that run on their own machines. A smaller model costs less to run and keeps data private.
Thinking Machines scored Inkling 2 Mini at 20.1 on the AA Index, a common AI benchmark. That mark trails huge models, but it is strong for a model this size.
What does this say about the people who build AI? Labs now race to shrink their models, as well as grow them. A model people can run on a laptop reaches far more builders. A model that needs a data center reaches fewer of them. Thinking Machines heard complaints about its last release and answered within weeks.