Thinking Machine Labs, founded by ex-OpenAI exec, launches first AI model Inkling

Thinking Machine Labs, the AI startup formed by ex-OpenAI CTO Mira Murati, has released its first AI model called Inkling. The startup says that Inkling is an open-weight model trained from scratch and can be heavily customised. Here are all the details.

In Short

• Inkling is an open-weight model that developers can download and modify
• It was trained from scratch by Thinking Machine Labs
• It matches Nvidia's Nemotron 3 model but at one-third of token use

Thinking Machines Lab, the artificial intelligence startup founded by former OpenAI CTO Mira Murati and other executives, has finally released its first in-house AI model, Inkling. The company says that Inkling was trained from scratch with multimodal capability, that is, it can understand text, code, audio, and images. The model is open-weight, with Thinking Machine Labs stating that developers and companies can download and modify it, unlike the flagship models from the likes of OpenAI and Anthropic.

Mira Murati wrote on X, “Our first model, Inkling. Trained from scratch, weights are open, fine-tunable on Tinker today.” Tinker is a tool developed by the startup to fine-tune AI models.

To give you some context, Thinking Machine Labs was formed in February last year. The startup received roughly $12 billion in seed funding at the time – the largest in history. And now, it finally has its own AI model.

What is Inkling?

According to Thinking Machine Labs, Inkling has 975 billion total parameters and 41 billion active parameters for a given task. The model was trained from scratch on 45 trillion tokens spanning text, images, audio and video, and can reason across all four. Though for now it produces only text outputs, including code, styled artefacts and structured data.

The model was designed to be open-weight, similar to models like DeepSeek v4 and Zai’s GLM 5.2. This means that you can download Inkling and run it locally, instead of having to use it via the cloud like GPT 5.6 or Claude Fable. This also makes open-weight models much cheaper to run. This could allow Inkling to get wider adoption as companies globally reconsider AI use due to rising costs.

Keep in mind that Thinking Machine Labs states that while Inkling is not the “strongest overall model available,” it is a combination of qualities that makes it a good “open-weights base for customisation.” That is, companies or developers can tailor Inkling according to their desired tasks and workflows.

Inkling is also tied closely to the company’s central pitch – AI systems that organisations can shape for themselves will perform better than centrally trained, one-size-fits-all models.

The startup says that Inkling is designed to give calibrated answers, including signalling uncertainty instead of guessing, and lets users adjust “thinking effort” to balance speed and performance.

How does Inkling compare to other models?

In a blog post, Thinking Machine Labs shared benchmark comparisons for various AI models. On Terminal Bench 2.1, Inkling matched Nvidia’s Nemotron 3 Ultra while using roughly a third of the tokens. Though it was behind the likes of GPT 5.6 and Claude Fable 5, as well as Kimi K2.6.

The company says that Inkling was pre-trained from scratch, but confirmed that early post-training data partly used outputs from other open-weight models, including Moonshot AI’s Kimi K2.5, before large-scale reinforcement learning took over.

The model was trained entirely on Nvidia GB300 NVL72 systems under a partnership announced in March.

Alongside Inkling, the company has also released a preview of Inkling-Small, a lighter-weight model with 12B active parameters. According to the company, this model should achieve strong power with “even lower cost and latency.” The company says that it plans to release more models in the Inkling family in the future.

The open-weight release of Inkling is inline with Thinking Machine Labs’s views on the future of AI. In a previous blog post, Thinking Machine Labs said that it believes AI should be decentralised rather than controlled by a few firms. That argument has also been echoed elsewhere. Microsoft chief executive Satya Nadella recently wrote that enterprises using proprietary models can end up paying both in subscription fees and by giving away business knowledge through prompts and corrections.

The release is the company’s first major public product after about a year and a half of building AI infrastructure largely out of view. Thinking Machine Labs is said to have about 200 employees at the moment.

- Ends