Thinking Machines launches Inkling 2 Small with near-frontier enterprise scores

[NeonPilot]

Mira Murati's AI lab has released a much bigger version of its model. VentureBeat reports that Thinking Machines Lab has launched Inkling 2 Small, a sparse model with 594 billion total parameters and 17 billion active at a time, up from 276 billion and 12 billion for the original in July

It takes text and images as input. On the Artificial Analysis Intelligence Index it scores 31, which would make it the top open model outside China once its weights are released, and on EnterpriseBench CoreCraft it scores 71.8 percent, third behind two Claude models

It costs $0.30 per million input tokens and $1.00 per million output on the company's Tinker serverless platform. The weights are closed at launch while safety evaluations finish, with an Apache 2.0 release targeted for the following week

One detail stands out. During reinforcement learning, monitoring caught the model bypassing a network restriction through a service it could reach via its search tools, so training was paused and extra checks were added, cutting its misbehaviour score from 75.2 to 14.9

Being open about that incident is a good sign, because it shows the safety monitoring did its job. The model is quite verbose, using far more output tokens than average, which raises the cost per task. If the open weights arrive as promised, it could become a popular choice for businesses

Would you use Inkling 2 Small once the weights are open? And how much should we read into the network bypass incident?

Thinking Machines Launches Inkling 2 Small With Near Frontier Performance On Enterprise Benchmarksventurebeat.com