Thinking Machines Lab, the startup founded by former OpenAI chief technology officer Mira Murati, released its first in-house AI model on Wednesday. Called Inkling, it is an open-weight system that outside developers can download, modify, and run on their own infrastructure — a deliberate contrast to the closed, subscription-based flagship models sold by OpenAI, Anthropic, and Google, according to the company.
What Inkling is
Inkling is a mixture-of-experts transformer with 975 billion total parameters, of which about 41 billion are active for any given task — a design that keeps very large models cheaper to run. According to its model card, each token is routed through 6 of 256 specialized experts plus 2 shared ones, across 66 layers, with a context window of up to one million tokens. The model was trained on a mix of text, images, audio, and video and can natively process all three input types, though it currently outputs text only. Weights are released under an Apache 2.0 license and are downloadable from Hugging Face, with hosted access also available through Tinker and third-party providers including Together AI, Fireworks, Modal, Databricks, and Baseten.
Thinking Machines is candid that Inkling is not a leaderboard leader. The company’s own materials describe it as “not the strongest overall model available today, open or closed,” positioning it instead as a customizable foundation. It also released Inkling-Small, a 276-billion-parameter preview variant with 12 billion active parameters that matches or beats the larger model on some benchmarks.
A different bet
The release tests the startup’s core thesis: that models organizations can adapt to their own data will outperform one-size-fits-all systems for many real-world tasks. The company has pointed to a collaboration with Bridgewater Associates in which fine-tuning an open model on proprietary financial expertise reportedly reached 84.7% on a financial-reasoning benchmark at a fraction of the cost of comparable proprietary alternatives. Tinker, the company’s fine-tuning platform, rather than the model itself, is the primary intended revenue source, and it is currently offering a limited-time discount for developers customizing Inkling.
Thinking Machines reached the market roughly nine months after founding, a far shorter runway than OpenAI or Anthropic took to ship their first flagship products. The model card flags standard limitations for a system of this kind, including a tendency to hallucinate, occasional lapses in following instructions, and a training-data cutoff, alongside possible demographic and linguistic bias.