Thinking Machines Lab, the AI startup founded by former OpenAI CTO Mira Murati, has released Inkling, its first open-weights artificial intelligence model.
On paper, Inkling is an ambitious system. It uses a Mixture-of-Experts transformer architecture with 975 billion parameters, although only 41 billion are active at any given time. The model can process text, images, audio and video, supports context windows of up to one million tokens and was trained on 45 trillion tokens.
The company has also introduced Inkling-Small, a lighter preview version that uses 12 billion active parameters.
However, the most interesting part of Inkling is not its size or technical specifications. Thinking Machines Lab has openly acknowledged that the model is not the most powerful option available, either among proprietary models or open-weights alternatives.
In benchmark tests such as SWE-Bench, Humanity’s Last Exam and SimpleQA, Inkling trails frontier models from OpenAI, Anthropic and Google. It also does not consistently outperform other open models. Nemotron 3 Ultra and GLM 5.2, for instance, deliver better results in some coding and reasoning evaluations.
So why release a model that does not sit at the top of the leaderboard?
Thinking Machines Lab believes Inkling’s real strength lies in customisation rather than raw intelligence. The model is available for fine-tuning through Tinker, the company’s AI training platform.
The idea is that a broad and well-rounded general-purpose model may be easier to adapt for specialised tasks than a highly optimised frontier model. Whether that strategy works will depend on how effectively developers can fine-tune Inkling for real-world applications, something that traditional benchmark scores cannot fully measure.
To demonstrate this capability, the company showed Inkling creating its own fine-tuning task, generating synthetic evaluation data and loading the updated weights back into a coding agent.
It was an impressive technical demonstration, but completing a controlled writing exercise does not necessarily prove that the model can handle complex, industry-specific fine-tuning requirements.
Inkling does offer several noteworthy features. It can reportedly assist with its own fine-tuning and allows developers to control how much computational effort it spends on reasoning. This gives teams the option to prioritise speed and lower costs or allow the model to spend more resources on producing higher-quality responses.
The model has also reportedly matched Nemotron 3 Ultra on Terminal Bench 2.1 while using only around one-third of the tokens. Thinking Machines Lab has additionally highlighted Inkling’s performance in Cognition’s censorship-resistance evaluation.
Inkling is available through Tinker and Hugging Face, as well as inference providers including Together AI, Fireworks, Databricks and Baseten.
For developers, the central question is not whether Inkling is the smartest model on the market. It clearly is not.
The real question is whether a customisable, capable general-purpose model can be adapted to perform a specific job more affordably than repeatedly paying for access to a more powerful but closed AI system.
That is the bet Thinking Machines Lab is making with Inkling, and its success could influence how developers across the open-weights AI ecosystem build specialised applications.
