A new open-weight AI model called Inkling is reshaping how developers think about the trade-off between AI reasoning power and practical cost. Rather than building the single strongest model available, Thinking Machines Lab designed Inkling to let developers control how much computational effort the model expends on each problem, making it possible to achieve strong performance at a fraction of the typical cost.

What Makes Inkling Different From Other Foundation Models?

Inkling is a Mixture-of-Experts transformer with 975 billion total parameters, though only 41 billion are active during inference, meaning the model uses less computing power than its total size might suggest. It was pretrained on 45 trillion tokens of text, images, audio, and video, and supports a context window of up to 1 million tokens, allowing it to process roughly 1 million words at once.

The model's defining feature is what Thinking Machines Lab calls "controllable thinking effort." Developers can adjust a setting from 0.2 to 0.99 to control how many tokens, or computational steps, the model generates when solving a problem. This matters because in real-world applications, cost and latency are often the limiting factors, not peak performance on academic benchmarks. On Terminal Bench 2.1, a benchmark for agentic coding tasks, Inkling matches the performance of Nemotron 3 Ultra using roughly one-third as many tokens.

Inkling also reasons natively over text, images, and audio without requiring separate encoding steps for vision and sound. The model uses an encoder-free architecture, processing audio as spectrograms and images as 40-by-40 pixel patches. This design choice aligns with Thinking Machines Lab's broader vision of building interaction models that enable real-time collaboration through voice and vision.

How Does Test-Time Compute Change What Developers Can Build?

Test-time compute, also called inference scaling, refers to the idea that AI models can improve their reasoning by spending more computational effort on a problem at the moment of use, rather than only during training. Inkling embodies this principle by letting developers sweep the effort setting across a range of benchmarks and observe the performance-to-efficiency curve. This approach acknowledges a fundamental reality: different use cases have different constraints.

For a model running millions of times as part of longer workflows, the ability to dial down computational effort without catastrophic performance loss is valuable. Thinking Machines Lab demonstrated this by showing Inkling's performance across three major benchmarks: Terminal Bench 2.1 for agentic coding, HLE for advanced reasoning, and IFBench for instruction following. On all three, Inkling reaches a given performance level at fewer tokens than competing open-weight models.

Steps to Evaluate and Fine-Tune Inkling for Your Use Case

• Test the Model Interactively: Thinking Machines Lab added the Inkling Playground to the Tinker console, a developer-facing interface for chatting with Inkling directly. This lets you get a qualitative feel for the model before committing to fine-tuning, since picking the right base model combines measurable benchmarks with subjective judgment about how a model behaves in practice.
• Measure Performance Across Your Effort Settings: Run your specific task or workflow at different effort levels, from 0.2 to 0.99, and plot the performance-to-token curve. This reveals the sweet spot where your model delivers acceptable accuracy at the lowest cost and latency for your application.
• Fine-Tune on Tinker: Inkling is available for fine-tuning on Tinker today. The model's broad training across agentic, reasoning, coding, instruction-following, factuality, vision, and audio tasks provides a strong foundation for customization across many domains.

Thinking Machines Lab also trained Inkling to run inside a variety of coding and agent harnesses, randomizing the tool set and schema during training to reduce sensitivity to any particular tool configuration. This flexibility matters for developers building agentic systems that need to adapt to different environments.

Why Multimodal Capabilities Matter for Foundation Models

Inkling's multimodal design reflects a shift in how foundation models are built. Rather than treating text, vision, and audio as separate problems, the model processes all three modalities jointly from the ground up. For audio, Inkling transcribes speech, follows spoken instructions, answers questions about recordings, and reasons over longer-form audio. On audio benchmarks like VoiceBench, MMAU, and AudioMC, it ranks among the strongest open-weight models.

For vision, Inkling accepts images as input and can describe visual content, answer questions, and perform in-depth reasoning based on visual information. It demonstrates strong performance on charts, diagrams, and mathematical visual reasoning tasks. During inference, the model can also leverage a Python tool to support image understanding through operations such as zooming and cropping, seamlessly integrating visual reasoning with code-based reasoning.

Thinking Machines Lab expects its multimodal capabilities to continue improving as the model and training pipeline expand in subsequent iterations. The company also released a preview of Inkling-Small, a lighter-weight variant with 12 billion active parameters, trained with a similar recipe, that achieves strong performance with even lower cost and latency.

What Does This Mean for the Broader AI Landscape?

Inkling represents a deliberate choice to prioritize customization and efficiency over raw benchmark dominance. Thinking Machines Lab acknowledged that Inkling is not the strongest overall model available today, open or closed. Instead, the combination of multimodal capabilities, efficient thinking, and availability for fine-tuning makes it a practical base for real-world applications.

The release signals a maturing market where developers increasingly care about the full cost curve of a model, not just its peak performance. As AI systems move from research labs into production environments, the ability to balance accuracy, speed, and cost becomes as important as raw capability. Inkling's controllable thinking effort feature makes this trade-off explicit and measurable, giving developers the tools to optimize for their specific constraints rather than accepting a one-size-fits-all model.