Thinking Machines unveils Inkling, an open-weight bet on customizable AI

Lead

Thinking Machines Lab, the startup founded by former OpenAI CTO Mira Murati, has made its first major public model release: Inkling. Unlike the flagship models from OpenAI, Anthropic and Google, Inkling is open-weight, meaning developers and companies can download it, run it and modify it directly. The release is less about claiming the top spot on leaderboards and more about making a strategic argument: enterprise AI should be shaped around an organization’s own expertise, not forced into a single general-purpose product.

Key points

• Open weights as the product strategy: Inkling is not being presented as a finished chatbot. Thinking Machines is framing it as a foundation that companies can customize, especially through its model-customization platform, Tinker.
• A large mixture-of-experts design: The model has 975 billion total parameters, but activates roughly 41 billion for a given task. That design aims to preserve scale while keeping inference faster and cheaper than using the full model each time.
• Multimodal training, text output for now: The company says Inkling was trained on 45 trillion tokens spanning text, image, audio and video, and can reason natively across those modalities. At launch, however, it outputs text, including code, styled artifacts and structured data.
• Calibrated answers and adjustable thinking: Inkling is designed to flag uncertainty instead of simply guessing. Users can also raise or lower “thinking effort” depending on whether they want more reasoning or more speed.
• Not the strongest model by its own admission: Thinking Machines explicitly says Inkling is not the strongest model available, open or closed. Its pitch is balanced performance plus adaptability.

Why it matters

Inkling lands at a moment when enterprises are questioning how much business knowledge they should hand to proprietary AI platforms. Closed frontier models offer convenience and strong general performance, but they also put prompts, corrections and workflow data inside systems controlled by another company. Thinking Machines is betting that many organizations will prefer models they can fine-tune, host and govern closer to their own operations.

That explains the role of Tinker. If Inkling’s weights are public, the model itself is not the main toll booth. Revenue has to come from customization, training, fine-tuning and potentially a hosting ecosystem around the model. This is a different economic model from metered API access: the value shifts from serving a fixed model to helping customers turn it into a domain-specific system.

There are trade-offs. Fine-tuning requires serious machine-learning talent, and safety responsibilities may shift toward the customer once the model is modified. The company also acknowledged that some early post-training data was generated with help from other open-weight models, including Moonshot AI’s Kimi K2.5, even though it says future models will use a fully self-contained post-training process. And with major Nvidia compute commitments in the background, questions remain about cost and funding.

Still, Inkling captures an important industry turn. Frontier closed models may remain essential for experimentation and high-value tasks, while more production workloads move toward private, open-weight or heavily customized systems. Thinking Machines’ bet is that the next enterprise AI question will not be only “Which model is strongest?” but “Which model can become ours?”

Source: TechCrunch AI