Inkling

Summary: Inkling is an open-weight, multimodal AI model developed by Thinking Machines Lab, released on July 15, 2026. Built using a Mixture-of-Experts (MoE) transformer architecture, it supports text, image, and audio understanding, reasoning, coding, and AI agent workflows. The model has 975 billion total parameters, with 41 billion active parameters, and supports a context window of up to 1 million tokens. Developers can customize and fine-tune Inkling through the company's Tinker platform or deploy its weights through supported AI infrastructure providers.

What is Inkling?

Inkling is a general-purpose open-weight foundation model created by Thinking Machines Lab. It is designed to support a broad range of AI applications, including conversational assistants, coding tools, multimodal applications, and autonomous AI agents. Unlike proprietary models that restrict access to their underlying weights, Inkling makes its full model weights available so developers and researchers can customize and integrate it into their own applications.

The model uses a Mixture-of-Experts transformer architecture with 975 billion total parameters and 41 billion active parameters. It was pretrained on approximately 45 trillion tokens covering text, images, audio, and video. Its architecture is designed to balance performance and computational efficiency by activating only a portion of its parameters for each token.

Inkling natively processes text, image, and audio inputs, making it suitable for multimodal AI applications. Its capabilities include reasoning, instruction following, coding, factual question answering, visual understanding, audio processing, and tool use. It also supports controllable thinking effort, allowing developers to balance response quality, processing time, and computational cost.

One of Inkling's major advantages is its customization capability. Through Tinker, Thinking Machines Lab's model-training platform, developers can fine-tune the model for specialized tasks and organizational requirements. The company also demonstrated Inkling fine-tuning itself to adopt a specific writing behavior, showcasing how its weights can be customized for targeted applications.

Inkling is designed for developers building AI-powered products rather than being only a consumer chatbot. Its weights are available through Hugging Face, while API and deployment options include Together AI, Fireworks, Modal, Databricks, and Baseten. It is released under the Apache 2.0 license, supporting research, customization, and commercial integration subject to the license terms.

Verifiable Stats

1. Tool Name: Inkling.

2. Developer: Thinking Machines Lab, Inc.

3. Launch Date: July 15, 2026.

4. Model Architecture: Sparse Mixture-of-Experts (MoE) transformer.

5. Total Parameters: 975 billion, with 41 billion active parameters.

6. Training Data: Approximately 45 trillion tokens of text, images, audio, and video.

7. Context Window: Up to 1 million tokens; Tinker currently offers 64K and 256K context options.

8. Input Modalities: Text, images, and audio.

9. Primary Capabilities: AI reasoning, coding, multimodal understanding, instruction following, and agentic tool use.

10. License: Apache 2.0 open-weight license.