What is Inkling?

Inkling is Thinking Machines’ first open-weights model, released under the Apache 2.0 license. It is a Mixture-of-Experts (MoE) transformer with 975 billion total parameters and 41 billion active parameters per forward pass. Inkling supports a context window of up to 1 million tokens and was pretrained on 45 trillion tokens spanning text, images, audio, and video. It reasons natively across these modalities and offers controllable thinking effort, allowing users to balance cost and performance. The model is available for fine-tuning on Tinker, and its full weights can be downloaded directly.

Who it's for

• Developers and AI engineers who want a strong, customizable open-weights base model they can fine-tune for specific use cases without vendor lock-in.
• Researchers and open-source contributors who need access to model weights under a permissive license (Apache 2.0) for experimentation, adaptation, or further training.
• Teams building multimodal applications that require native reasoning across text, images, and audio with efficient, controllable inference costs.

Key features

975B MoE with 41B active parameters

Inkling uses a Mixture-of-Experts architecture that activates only a fraction of its total parameters per token. This design delivers the capability of a large model while keeping inference compute efficient and latency manageable.

1M token context window

The model supports up to 1 million tokens of context, making it suitable for long-document analysis, extended conversations, and tasks that require reasoning over large amounts of information in a single pass.

Native multimodal reasoning

Inkling was pretrained on text, images, audio, and video, and it reasons across these modalities natively. It does not rely on separate encoders or pipelines for different input types.

Controllable thinking effort

Users can adjust how much computation the model spends on reasoning before generating an answer. This allows fine-grained control over the trade-off between response quality and cost or latency.

What stands out

Inkling is not the strongest overall model available today — instead, a combination of qualities makes it a good open-weights base for customization.

This honest positioning sets Inkling apart. Rather than chasing top benchmark scores, Thinking Machines focused on building a balanced, flexible foundation model that excels as a starting point for fine-tuning. The model is available on Tinker for immediate customization, and the team demonstrated this by having Inkling fine-tune itself — writing its own training job, running it, and evaluating the result — all within the Tinker platform.

Worth checking out if…

You are looking for a permissively licensed, open-weights model that balances multimodal capability with efficient inference, and you value the ability to fine-tune and customize the model for your own use cases. Inkling is especially worth exploring if you want to experiment with controllable reasoning effort or need a model that can handle long contexts across text, images, and audio.