Thinking Machines Lab is releasing Inkling, an open-weights Mixture-of-Experts model with 975 billion parameters, 41 billion of which are active. The model has a context window of 1 million tokens and processes text, images, audio, and video. Fine-tuning can be done immediately via Tinker.
Inkling was trained on 45 quadrillion tokens of text, images, audio, and video, and is available entirely as an open-weights model. In addition to the main model, the company is sharing a preview of Inkling-Small, a lighter variant with 12 billion active parameters that was trained in a similar manner but runs at lower cost and with lower latency.
“Inkling is not the strongest overall model available today, open or closed,” the company states. Instead, it aims for a combination of features: multimodal capabilities, efficient and configurable reasoning, and immediate availability on Tinker for fine-tuning.
Tip: Former OpenAI CTO Murati launches startup Thinking Machines Lab
Multimodal and agentic
Inkling is trained on agentic tasks, reasoning, coding, following instructions, and vision and audio tasks. The “thinking effort” is adjustable, allowing users to balance cost and performance. For agentic work, the company trained the model within varying coding and agent harnesses, using random tools to limit dependence on any single specific set.
To demonstrate the benefits of customization, the company had Inkling fine-tune itself: using Tinker, the model wrote its own fine-tuning job, executed it, and evaluated the result. Developers can also chat with the model in the Inkling Playground within the Tinker console.
According to the company, Inkling is just the beginning: it is the first release in a family of models on which it plans to build further. The model is available for fine-tuning on Tinker starting today.