Thinking Machines Lab has introduced Inkling 2 Small, the first release in its second model generation. Its announcement says the model can take image inputs and vary its reasoning effort, with initial access through the company’s Tinker platform.
What Inkling 2 Small can process
The company describes Inkling 2 Small as a mixture-of-experts model with 594 billion total parameters and 17 billion active parameters. The active number describes the subset used for computation and sits alongside the full parameter count. Inkling 2 Small combines a vision encoder with text processing so an agent can use visual feedback while refining a website, game or design.
Thinking Machines Lab says it trained the model on long-horizon coding tasks using multiple agent harnesses. Some training tasks used a grader that checked the process as well as the result. That distinction matters to developers evaluating whether the model can test and revise a change over several steps.
How to try it
The announcement says users can try the model in Tinker Playground. Tinker also offers sampling and fine-tuning. The company describes the initial weights as closed and says it intends to release the full model weights under Apache 2.0 after external safety testing and red-teaming. Checkpoint downloads are tied to that release.
How to read the reported scores
The announcement reports a score of 31 on the Artificial Analysis Intelligence Index. The company also gives a score of 81.6% on MMMU Pro with the Standard 10 setup. Those numbers describe different evaluation conditions. Teams considering the model for a visual coding workflow can try a representative task in the playground and check whether the model’s revisions preserve details from the original image.
