What Thinking Machines Lab's new open model Inkling means for you

Thinking Machines Lab, the start-up founded by former OpenAI chief technology officer Mira Murati, has released its first open-weights model, called Inkling. Open-weights means the underlying model files are published for anyone to download and build on, rather than locked behind a paid API.

Inkling is a mixture-of-experts model, a design where only a small portion of the model's total knowledge is "switched on" for each request, making it cheaper to run than its full size suggests. It has 975 billion parameters in total (parameters are the adjustable settings that store what a model has learned), but only 41 billion are active at any one time. It was trained on 45 trillion tokens of text, images, audio and video, and can handle up to 1 million tokens of context at once, meaning it can read and reason over very long documents or conversations in one go.

Thinking Machines also released a smaller version, Inkling-Small, with 12 billion active parameters; its full weights will follow once testing is complete.

On the lab's own benchmark tests, Inkling scored 63.8% on Terminal Bench 2.1, 97.1% on AIME 2026 (a maths competition test), 87.2% on GPQA Diamond (graduate-level science questions), and 77.6% on SWE-bench Verified (real coding tasks). The lab says these reasoning scores come at roughly a third of the token cost of rival models, a measure of how much computing power, and therefore money, a task uses.

Crucially, Thinking Machines is not claiming Inkling is the best model available. As the lab put it in its launch post: "Inkling is not the strongest overall model available today, open or closed. Instead, a combination of qualities makes it a good open-weights base for customization."

The release lands the same week as Moonshot's 2.8-trillion-parameter Kimi K3, part of a wider push by both US and Chinese labs to close the gap with the top closed models.

If you're not a developer, Inkling won't change much day to day. It's not a consumer chatbot app. Its value is to companies and developers who want to fine-tune, or further train, a capable AI model for their own specific use, such as customer support, coding tools or industry-specific assistants, without relying on a closed model from a big provider.

Developers can access Inkling's weights on Hugging Face, or fine-tune it through Thinking Machines' Tinker platform, currently offered at a 50% discount for a limited time. It's also available via API through Together AI, Fireworks, Modal, Databricks and Baseten.