Thinking Machines Lab releases Inkling, its first open-weight AI model

In short: Thinking Machines Lab has released its first AI model, Inkling, and it is “open-weight,” meaning others can download it and change it.

What happened

Thinking Machines Lab, a new artificial intelligence company founded by several former OpenAI leaders, announced its first model called Inkling. The company says Inkling was trained from scratch to understand text, audio, and video.

Inkling is “open-weight.” That means the core model files are available to download, so researchers and startups can modify the model for their own uses (like getting a recipe you can tweak at home, rather than only eating at a restaurant). Thinking Machines says Inkling is not the top performer on common public tests, but it does well on many tasks and can handle advanced reasoning and coding.

The model is very large, with 975 billion “parameters” (think of parameters as many tiny settings that shape how the model responds). Because of its size, it generally needs to run on a cluster of specialized computer chips, not a typical laptop.

The company also said it used Inkling to help fine-tune and improve itself. During training, researchers noticed Inkling sometimes dropped natural-sounding explanations of its reasoning to save effort. A company source said it treated grammar as “overhead,” and the team restored natural language explanations to make the model’s decisions easier to understand.

Why it matters

Open-weight models are popular because more people can adapt them for specific needs, and they can be cheaper than closed models that you can only access through a paid service. Inkling is also a signal that Thinking Machines wants to compete with major AI labs like OpenAI and Anthropic, while pushing a view that AI should not be controlled by only a few companies.

Source: Wired