Thinking Machines Lab has released Inkling, an open-weight foundation model designed to process text, images, audio, and video simultaneously, making it one of the first multimodal AI systems available for customization by developers and researchers. The model contains 975 billion total parameters, with 41 billion active during inference, and was pretrained on 45 trillion tokens of multimodal data. Unlike closed proprietary models, Inkling's full weights are publicly available, allowing anyone to fine-tune it for specialized tasks.

What Makes Inkling Different From Other Open-Source AI Models?

Inkling stands out for its efficiency and multimodal design. The model reaches the same performance level as competitors like Nemotron 3 Ultra while using roughly one-third as many computational tokens, according to benchmarks on Terminal Bench 2.1. This efficiency matters significantly for real-world applications where cost and speed are critical constraints. The model also supports a context window of up to 1 million tokens, meaning it can process roughly 100,000 words at once, enabling it to handle longer documents and conversations without losing context.

Thinking Machines Lab designed Inkling as a generalist model rather than optimizing narrowly for a single task. The company trained it across agentic reasoning, coding, instruction-following, factuality, vision, and audio tasks, making it flexible enough to adapt to different workflows. This breadth is intentional; developers fine-tuning models for specialized applications need systems that can serve as strong foundations across multiple domains.

How Does Inkling Process Audio and Visual Information?

Inkling's multimodal capabilities represent a significant technical achievement. For audio, the model accepts dMel spectrograms as input and can transcribe speech, follow spoken instructions, answer questions about recordings, and reason over longer-form audio content. On audio benchmarks including VoiceBench, MMAU, and AudioMC, Inkling ranks among the strongest open-weight audio models available. For vision, the model accepts images as input and can describe visual content, answer questions, and perform reasoning based on visual information. It demonstrates particularly strong performance on charts, diagrams, and mathematical visual reasoning tasks.

The architecture uses an encoder-free design for both audio and vision inputs, meaning the model processes these modalities directly without separate preprocessing layers. Images are encoded as patches of 40 by 40 pixels using a lightweight embedding approach, while audio signals are transformed via a light-weight embedding layer and processed jointly with text tokens. This unified approach allows the model to reason across modalities seamlessly.

How to Get Started With Inkling for Your Organization

• Access the Model: Inkling is available for fine-tuning on Tinker, Thinking Machines Lab's platform, allowing developers to customize the model for domain-specific applications without building from scratch.
• Test Before Committing: The company added an Inkling Playground interface within the Tinker console, a developer-facing tool for chatting with Inkling to evaluate its capabilities and feel before deciding whether to fine-tune it for your use case.
• Choose the Right Size: Alongside the full 975-billion-parameter model, Thinking Machines Lab is previewing Inkling-Small, a lighter-weight variant with 12 billion active parameters that achieves strong performance with lower cost and latency for resource-constrained environments.

What Does Controllable Thinking Effort Mean for Developers?

One of Inkling's distinctive features is controllable thinking effort, which allows developers to adjust how much computational work the model expends on a given problem. By sweeping the effort setting from 0.2 to 0.99, developers can trace the model's performance against the number of tokens generated, essentially trading off accuracy for speed and cost. This flexibility is crucial for real-world applications where latency constraints are binding. For example, a customer service chatbot might use low effort for simple queries but higher effort for complex reasoning tasks. Inkling reaches a given performance level at fewer tokens than competing models, making it more efficient across the board.

This capability addresses a fundamental tension in AI deployment. Test-time scaling and problem-solving are core to every model's capability, but that capacity is difficult to capture with a single benchmark number. Developers fine-tuning models for specialized tasks care as much about efficiency as about maximum performance on public benchmarks. Cost and latency are often the binding constraints in production systems, and low latency is particularly crucial for enabling real-time collaboration and iterative improvement.

Why Did Thinking Machines Design Inkling as a Multimodal Model?

Thinking Machines Lab's primary goal for Inkling is to serve as the background reasoning engine in the company's interaction models system, which enables users to collaborate naturally using voice and vision in real time. This required a model natively trained for broad multimodal capabilities from the ground up. Rather than bolting on audio and vision capabilities after the fact, the company trained the multimodal components from scratch on general-domain data, ensuring seamless integration across modalities.

The company acknowledges that Inkling is not the strongest overall model available today, whether open or closed. Instead, a combination of qualities makes it a good open-weights base for customization: multimodal capabilities, efficient thinking, and availability on Tinker for fine-tuning. This honest positioning reflects the reality that different use cases require different trade-offs. A model optimized purely for benchmark performance might not be the best choice for a developer who needs to fine-tune it for a specific domain or who faces latency constraints in production.

What's Next for Thinking Machines Lab?

Inkling is positioned as the first release in a family of models that Thinking Machines Lab will continue to build on. The company expects its multimodal capabilities to continue improving as the model and training pipeline are expanded in subsequent iterations. The release of Inkling-Small alongside the full model suggests the company is committed to serving developers across different computational budgets and latency requirements.

The broader significance of Inkling lies in its approach to open-weight model development. By releasing full weights and making fine-tuning accessible through Tinker, Thinking Machines Lab is democratizing access to multimodal AI capabilities. This contrasts with closed proprietary models where customization is either unavailable or restricted to enterprise customers. For researchers, startups, and organizations building specialized AI applications, open-weight multimodal models like Inkling represent a meaningful shift in what's possible without massive budgets or partnerships with major AI labs.