Inkling

• Open-weight 975B-parameter Mixture-of-Experts multimodal model
• Native reasoning across text, images, and audio inputs
• Up to 1 million-token context window
• Controllable thinking effort for balancing performance, cost, and latency
• Agentic coding and tool-use capabilities
• Image understanding for visual questions, charts, diagrams, and documents
• Audio understanding for speech, recordings, and longer-form audio
• Fine-tuning support through the Tinker training platform
• Open model weights available in standard and quantized checkpoint formats
• API availability through Tinker and supported third-party inference providers

Inkling is an open-weight multimodal AI model developed by Thinking Machines Lab for developers and researchers building customizable AI applications. Released on July 15, 2026, Inkling is designed as a general-purpose model capable of reasoning across text, images, and audio while supporting coding, tool use, and specialized fine-tuning workflows.

The primary Inkling model uses a Mixture-of-Experts architecture with 975 billion total parameters and 41 billion active parameters. It supports a context window of up to 1 million tokens, although available context lengths can vary depending on the deployment environment. The model accepts text, image, and audio inputs and produces text outputs.

A key capability is controllable thinking effort. Developers can adjust the model’s reasoning effort to balance performance, token usage, cost, and latency for different applications. This can be useful when the same model needs to handle both lightweight interactions and more demanding reasoning tasks.

Inkling also supports agentic coding and tool-use workflows. It can write code, interact with tools, and work as part of applications that use external systems. Its multimodal capabilities allow developers to build applications involving image understanding, document analysis, visual question answering, speech processing, and audio reasoning.

The model is available with open weights under the Apache 2.0 license, allowing developers and researchers to download and customize the released checkpoints. Inkling can also be fine-tuned through Tinker, Thinking Machines Lab’s model training platform. Developers can use the Inkling Playground in Tinker to interact with the model before fine-tuning.

Inkling is available through Tinker and several third-party inference providers. Running the model directly requires substantial GPU resources, while quantized checkpoints can reduce the hardware requirements.

Inkling is primarily intended for AI developers, researchers, organizations, and technical teams that need an open-weight multimodal foundation model that can be adapted to specialized applications.

Frequently Asked Questions

What is Inkling?

Inkling is an open-weight multimodal AI model from Thinking Machines Lab designed for reasoning, coding, tool use, image understanding, and audio processing.

When was Inkling released?

Inkling was released on July 15, 2026.

Is Inkling free?

Inkling’s model weights are openly available under the Apache 2.0 license, while API and fine-tuning usage through Tinker is paid according to usage.

What inputs does Inkling support?

Inkling supports text, image, and audio inputs and generates text outputs.

How large is the Inkling model?

Inkling has 975 billion total parameters, with 41 billion active parameters per token in its Mixture-of-Experts architecture.

Can Inkling be fine-tuned?

Yes, Inkling can be fine-tuned through Tinker, allowing developers to adapt the model to specialized tasks and domains.

Does Inkling support long contexts?

Yes, Inkling supports a context window of up to 1 million tokens, with shorter context limits available in some Tinker configurations.

Who is Inkling designed for?

Inkling is designed for developers, researchers, AI application builders, and organizations creating customizable multimodal AI systems.

Share Your Experience: