594B-Parameter Model Dubbed 'Small'

Thinking Machines Lab has officially released Inkling 2 Small.

Despite carrying the name 'Small', the model has a total of 594 billion parameters. Built on a Mixture of Experts (MoE) architecture, it has only 17 billion activated parameters per inference pass, and scores 31 points on the Artificial Analysis Intelligence Index. It currently operates on the Tinker platform, with Apache 2.0 licensed weight releases planned.

First, some background: Thinking Machines Lab is the company founded by Mira Murati after she departed OpenAI. Tinker is their developer-facing API platform, designed primarily for model fine-tuning and inference workloads. Adding Inkling 2 Small to the platform effectively gives Tinker a flagship-tier model offering.

The decision to label a 594B model 'Small' is notable on its own. Per standard MoE industry conventions, total parameters define a model's knowledge capacity, while activated parameters determine inference cost. The 17B activated count means the computational load for a single forward pass is roughly equivalent to a 20B-class dense model, while the model holds 594B worth of stored knowledge. The math checks out: inference costs stay manageable, while the model's capacity sits in a completely different league.

A 31-point score on the Intelligence Index does not land it at the top of the leaderboard. But when you compare a 17B-activated model against alternatives that routinely have hundreds of billions of activated parameters, its efficiency is far more compelling than its raw benchmark score. The Artificial Analysis Intelligence Index is a third-party comprehensive evaluation: the score should not be treated as the single only performance standard, but it is more than sufficient as a cross-model reference point.

There is one extra detail: the Apache 2.0 licensed weights are only 'in the planning phase' for now. 'Planned' does not equal already released, and no timeline has been announced. Users looking to self-host the model will have to wait. The only available way to use it right now is via the Tinker API, for both fine-tuning and inference.

Model naming in the AI industry is getting more and more abstract. If a ~600B model is called 'Small', how massive would its 'Large' variant be? We will leave that question open for now.