Thinking Machines Lab releases Inkling open-weights model

Thinking Machines Lab has released Inkling, its first model trained from scratch, with the full weights available for developers to download and customize. The company announced the model on July 15, positioning it as a flexible multimodal foundation rather than a closed assistant that can only be accessed through one provider.

Inkling is a Mixture-of-Experts transformer with 975 billion parameters in total and 41 billion active for each token. It was pretrained on 45 trillion tokens spanning text, images, audio and video, and supports a context window of up to one million tokens. The model can reason across text, images and audio, use tools, write code and vary how much computation it spends on thinking. Thinking Machines says the adjustable effort level lets developers trade performance for lower token use, cost and latency.

The company is unusually direct about the model’s position: Inkling is not presented as the strongest open or closed model overall. Its value is the combination of broad capabilities, available weights and support for fine-tuning. On the company’s own published benchmarks, it is competitive with other open-weights models across reasoning, coding, tool use, vision and audio, but results vary considerably by task and many leading closed models still score higher.

Customization is the central product idea. Inkling is available through Thinking Machines’ Tinker platform, where users can fine-tune it and test it in a playground. The company demonstrated that workflow by asking Inkling to write and run its own fine-tuning job, evaluate the result and load the updated weights. Full model weights are also available on Hugging Face, while API deployment is supported by providers including Together AI, Fireworks, Modal, Databricks and Baseten.

For makers, this creates a new base model for specialized assistants, coding agents and multimodal applications without depending entirely on a proprietary frontier API. Open weights offer more control over deployment, adaptation and testing, although they do not automatically make a model cheap or simple to operate. A model of this scale still demands substantial infrastructure.

For businesses, Inkling is another sign that the AI market is shifting from choosing one general chatbot toward adapting models for particular workflows and internal knowledge. Thinking Machines is betting that customization, efficiency and ecosystem support can matter as much as winning every benchmark.