Thinking Machines amps up its bet against one-size-fits-all AI with its first open model, Inkling

Summary

Thinking Machines Lab has released Inkling, an open-weights Mixture of Experts (MoE) model with 975 billion parameters. The model features 41 billion active parameters. The company positions Inkling not as a finished product, but as a customizable starting point for enterprises. It also explicitly acknowledges that it is not currently the most powerful model available.

Why it matters

This move poses a direct structural challenge to the business models of OpenAI, Anthropic, and Google, which are built on closed models and metered access. The company is betting that enterprise fine-tuning will outperform general-purpose frontier models. Satya Nadella of Microsoft and Clement Delangue of Hugging Face share this view. The growing industry consensus that production AI is shifting toward open and private solutions could fundamentally change where AI revenue flows.

Highlights

• A total of 975 billion parameters, with approximately 41 billion active parameters per task thanks to the MoE architecture. The model was trained on 45 trillion tokens, including text, images, audio, and video.
• The Bridgewater Associates project scored 84.7 points in financial reasoning. The cost is approximately 1/14th of top-tier closed models. However, the evaluation was self-reported by the two companies involved.
• The revenue model relies more on the Tinker fine-tuning platform than on the model itself. Due to the open weights, no one is obligated to pay Thinking Machines.

Critical take

The core value proposition places the burden of fine-tuning, security, and quality on customers who may lack significant machine learning capabilities. On the other hand, the headline Bridgewater comparison was self-evaluated by the companies involved. Using competitor model outputs (Kimi K2.5) as post-training data raises distillation concerns, as acknowledged by the company. This suggests that the 'developed from scratch' claim is more nuanced than implied in the marketing copy.