Inkling is an open-weights large language model created by the company Thinking Machines, first released on July 15, 2026, under an Apache 2.0 license. It accepts text, images, and audio as input and can perform tasks such as chat, software development, multilingual applications, and agentic systems. As an open-weights model, its parameters are publicly available, allowing developers to inspect and fine-tune it for specialized uses.
The model is notable for its native reasoning across multiple modalities and its adjustable reasoning budget, which lets users trade between speed and thoroughness. It was designed for real-world efficiency and censorship resistance, with creators recommending external moderation tools for safety.
Training
Inkling is the first large language model released by Thinking Machines. It was trained on 45 trillion tokens comprising text, image, audio, and video data, enabling it to reason natively from any of these four formats. However, at the time of release, the model could only generate text, including source code and structured data, not images or audio.
The training process likely leveraged deep learning techniques common to modern Transformer (architecture) architectures, though specific details on architecture and optimization have not been publicly disclosed.
Features
Inkling was designed to be adaptable and efficient in production environments. A key feature is its programmable reasoning budget, which can be adjusted on a scale from 0.2 to 0.99. This controls the amount of "thought" the model performs before generating an output, allowing developers to optimize the cost/performance trade-off for each task. Lower budgets yield faster responses, while higher budgets produce more refined answers.
Unlike most other large language models, including other open-weights variants, Inkling was created with a focus on resistance to censorship. It provides direct answers on politically sensitive or censored topics. The creators advise using external content moderation tools because internal safety evaluation may not always prevent harmful responses. This philosophical stance distinguishes it in the generative AI landscape.
For customization, developers can fine-tune the model through the Tinker platform, also from Thinking Machines, and deploy it via third-party providers.
Technical Information
Inkling supports a context window of up to 1,048,576 tokens, enabling processing of very long documents or conversations. It contains 975 billion parameters, though only 41 billion are active at any given time due to its Mixture of experts design. Inputs are routed to the most relevant expert modules, improving computational efficiency.
The model can be run using common inference libraries such as llama.cpp. Minimal hardware requirements include at least 2 TB of VRAM, but a quantized version reduces this to approximately 600 GB, making deployment more feasible for organizations with smaller infrastructure.
Deployment and Ecosystem
Inkling's open-weights nature means it can be deployed on various cloud platforms, including Amazon Web Services, Azure, Google Cloud, and Oracle Cloud, or on-premises with suitable hardware. It is designed to integrate with agentic systems, where it can act as a reasoning and decision-making component. Its multilingual capabilities make it suitable for international applications.
Thinking Machines positions Inkling as a foundation model for building specialized AI tools, with the Tinker platform serving as the primary fine-tuning interface. As of its release, Inkling represents a notable contribution to the open-source AI community, offering a blend of performance, flexibility, and philosophical openness absent from many commercial counterparts.