Inkling: The Open‑Weights AI Model That Lets You Control the Brainpower

Inkling, the latest release from Thinking Machines Lab, is turning heads by offering a 975‑billion‑parameter Mixture‑of‑Experts (MoE) model with only 41 billion active parameters at any moment. The model ships with Apache 2.0‑licensed weights, making it the first truly open‑source, multimodal transformer that can process text, image, and audio in one pass. Developers and researchers now have a new foundation for building applications that need large‑scale reasoning without the cost of proprietary APIs.

Why Inkling is a Game Changer

Inkling is not marketed as the most powerful model available, but it excels in areas that matter for custom solutions: controllable thinking effort and extreme context length. The 1 million‑token window lets you feed entire books or datasets into a single inference run, reducing latency for long‑form content analysis.

Key Features at a Glance

• 975B‑parameter MoE architecture with 41B active experts – dynamic scaling for cost‑effective inference.
• Native multimodal input: text, image, and audio handled natively without separate encoders.
• 1M‑token context window – the longest in any open‑weights model to date.
• Fully open weights under Apache 2.0 – copy, modify, and redistribute with no restrictions.
• Controllable thinking effort – adjust the number of active experts to balance speed and accuracy.

Open‑Weights and Customization

One of the biggest hurdles in AI deployment is lock‑in. With Inkling, you can run the model on-premises or in any cloud environment, fine‑tune on proprietary data, and tweak the expert activation logic to fit your workload. The mixture‑of‑experts design means you don't pay for the full 975 billion parameters unless you need them.

What This Means for Developers

For data scientists, Inkling offers a playground to experiment with extreme context and multimodal fusion. Product teams can prototype voice‑enabled image analysis or text‑heavy recommendation engines without vendor lock‑in. The model’s open licensing also paves the way for community‑driven safety and bias audits.

Industry Reactions

Tech analysts note that Inkling’s release signals a shift toward democratizing large‑scale AI. While closed models from giants like OpenAI and Google still dominate in raw societies, open‑weights MoE models like Inkling provide a viable alternative for niche use cases that need transparency and flexibility.

Looking Ahead

Thinking Machines Lab plans to iterate on Inkling’s architecture, adding more experts and refining the expert‑selection algorithm_lvl. They also aim to expand the open‑source community contributions through a public repository and a dedicated forum for fine‑tuning recipes.

Ready to explore the future of customizable AI? Dive into Inkling’s documentation, join the community, and start building