Thinking Machines Lab Introduces Open-Weights Model - Inkling

Thinking Machines Lab, an artificial intelligence research and product company, has introduced Inkling - a Mixture-of-Experts transformer with 975B total parameters, 41B active. It supports a context window of up to 1M tokens. It was pretrained on 45 trillion tokens of text, images, audio, and video. It is the first in a family of models of different sizes: alongside it, the company is sharing a preview of Inkling-Small, a lighter-weight model with 12B active parameters, trained with a similar recipe, that achieves strong performance with even lower cost and latency.

The mission is to build AI that extends human will and judgment. A platform has been developed to enable model customization, alongside the preview of an AI system designed for interactive collaboration and the publication of novel research. Advancing this mission, a new model trained from scratch is being released with its full weights available, enabling broad adaptation and customization.

Inkling is a multimodal foundation model that reasons across text, images, and audio while balancing performance with efficient, controllable thinking. Designed for customization, it combines open weights, strong cross-domain capabilities, and availability on Tinker for fine-tuning. This release marks the first model in the Inkling family, with future iterations to follow.

To make model customization more accessible, Inkling is now available for fine-tuning on Tinker. The platform also introduces the Inkling Playground, a developer interface that enables interactive testing and exploration before fine-tuning.

Capabilities.

Built for real-world applications, Inkling combines broad multimodal capabilities with the flexibility to improve through fine-tuning. It is designed as a general-purpose foundation model, delivering strong performance across reasoning, coding, agentic workflows, instruction following, vision, audio, and factual accuracy.

The model supports advanced agentic coding and tool use, enabling complex workflows such as web application development, iterative software refinement, and artifact generation. Controllable thinking effort allows developers to balance performance, latency, and cost for different use cases.

Native multimodal reasoning across text, images, and audio enables capabilities ranging from speech understanding and visual reasoning to interactive collaboration. Inkling also emphasizes trustworthy outputs through calibrated reasoning, strong instruction following, and factual reliability, while built-in safety safeguards help ensure responsible deployment. Together, these capabilities position Inkling as a flexible open-weights foundation model for a wide range of AI applications.

The Making of Inkling.

Inkling is built on a Mixture-of-Experts (MoE) Transformer architecture optimized for efficiency, long-context understanding, and scalable performance. The model was pre-trained on 45 trillion multimodal tokens, including text, images, audio, and video, and further refined through supervised fine-tuning and large-scale reinforcement learning.

Training emphasized broad capabilities across reasoning, coding, tool use, multimodal understanding, and safety. Reinforcement learning at scale significantly improved problem-solving performance while enabling controllable thinking effort, allowing the model to balance reasoning quality with computational efficiency. Together, these architectural and training advances establish Inkling as a scalable foundation for future generations of open-weight AI models.

Inkling Availability.

Inkling is available on Tinker with 64K and 256K token context options, alongside a limited-time discounted offering. The platform includes updated fine-tuning support, new cookbook recipes highlighting multimodal capabilities, and tools for reliable sampling and post-training with reasoning, tool calls, and multimodal inputs.

The Inkling Playground enables developers to explore the model through an interactive chat interface with integrated web search before fine-tuning. Inkling checkpoints are also supported across major inference and deployment platforms, with integrations spanning APIs, open-source inference frameworks, and model tooling ecosystems.

Full Inkling weights are available on Hugging Face, including optimized NVFP4 checkpoints for efficient inference on NVIDIA Blackwell systems.