Thinking Machines is not claiming that its first AI model is the smartest on the market. It is betting that businesses may care more about control.
The San Francisco startup founded by former OpenAI CTO Mira Murati unveiled Inkling, an open-weight model that developers and enterprises can download, modify, and fine-tune for specialized uses. Unlike ChatGPT or Claude, Inkling is positioned less as a finished assistant and more as a foundation for companies building their own AI systems.
That strategy puts Thinking Machines into the growing contest between proprietary AI platforms and models businesses can customize and deploy on their own terms.
A different strategy
Unlike leading proprietary models from OpenAI, Anthropic, and Google, Inkling is designed as a foundation for customization rather than as a finished chatbot.
The model is a mixture-of-experts transformer with 975 billion total parameters and 41 billion active parameters per task. It was trained on 45 trillion tokens spanning text, images, audio, and video, supports context windows of up to one million tokens, and can reason across text, images, and audio while currently generating text outputs.
Thinking Machines acknowledged that Inkling does not lead the AI industry on raw benchmark performance.
"Inkling is not the strongest overall model available today, open or closed," the company said. Instead, it said the model was built to balance multimodal capabilities, efficient reasoning, and adaptability through fine-tuning on its Tinker platform.
Alongside the flagship model, the company also previewed Inkling-Small, a lower-cost version with 12 billion active parameters that will be released after testing.
Enterprise focus
Thinking Machines is positioning Inkling for organizations that want to build specialized AI systems using their own knowledge and data.
The company has made the model available for fine-tuning through Tinker while also publishing its weights on Hugging Face. It said Inkling is available on Tinker with 64,000- and 256,000-token context options and is launching with a limited-time discount.
Inkling enters a market where enterprises are weighing proprietary AI services against open-weight models that offer greater control over deployment and customization. The company also said Inkling allows developers to adjust its "thinking effort," enabling them to trade reasoning performance for lower cost and latency depending on their workloads.
Industry implications
Inkling's launch reflects a growing divide in the AI market between companies building closed, subscription-based models and those promoting open-weight systems that enterprises can own and customize.
That approach could appeal to businesses concerned about infrastructure costs, intellectual property, and the deployment of AI in their own environments. However, it also places more responsibility on customers to safely fine-tune and govern their models, a process that typically requires significant machine learning expertise.
Thinking Machines also disclosed that while Inkling was pretrained from scratch, some early post-training data was generated using existing open-weight models, including Moonshot AI's Kimi K2.5, before reinforcement learning became the primary training method.
Looking ahead
Inkling marks Thinking Machines' first major product test after emerging from stealth with multibillion-dollar backing. Rather than competing head-to-head with frontier models on benchmark scores, the startup is betting that flexibility and customization will become more valuable to enterprise customers over time.
Whether that strategy succeeds will depend less on Inkling's standalone performance than on how many organizations embrace Tinker as a platform for building specialized AI systems. If enterprises continue shifting toward models they can control, Thinking Machines could carve out a distinct position in a market currently dominated by proprietary AI providers.
Also read: Anthropic’s Claude for Teachers gives verified U.S. K-12 educators free access to lesson planning, classroom tools, and AI training.