Thinking Machines launches Inkling, challenging one-size-fits-all AI with open model
Thinking Machines Lab has released Inkling, an open-weight AI model designed for enterprise customization rather than general-purpose chatbot utility. The startup is betting that organizations will prefer adaptable infrastructure over the proprietary, closed-source models current
Thinking Machines Lab, the startup founded by former OpenAI CTO Mira Murati, released its first in-house AI model, Inkling, on Wednesday. According to TechCrunch, the launch concludes a year and a half of private infrastructure development. Unlike the proprietary systems offered by OpenAI, Anthropic, or Google, Inkling is an open-weight model, allowing developers to download and modify the software directly. The model is a mixture-of-experts system containing 975 billion total parameters, though it utilizes only 41 billion parameters per task. This design architecture is intended to maintain speed and cost-efficiency while handling complex reasoning across text, image, audio, and video inputs.
While the model supports multimodal inputs, its current outputs are restricted to text, code, structured data, and styled artifacts. Thinking Machines explicitly stated in a blog post that Inkling is not the strongest overall model currently available. Instead, the company is positioning it as a starting point for organizations to fine-tune using Tinker, its proprietary customization platform. The company claims that on specific coding benchmarks, Inkling achieves equivalent performance to Nvidia’s Nemotron 3 Ultra while consuming one-third of the tokens. Users can also adjust thinking effort settings to balance speed against accuracy, a feature intended to give enterprises more control over model behavior.
The release highlights a strategic divergence from the industry standard of general-purpose chatbots. Thinking Machines argues that centrally trained, static models underperform when compared to systems that organizations can adapt to their specific expertise. This approach aligns with broader industry critiques, such as those from Microsoft CEO Satya Nadella, who recently warned that enterprises using proprietary models may face hidden costs through subscription fees and the loss of proprietary data used to train future model versions. Thinking Machines intends to generate revenue through its Tinker platform rather than selling metered access, effectively shifting the business model from service-based to infrastructure-based.
Regarding training data, the company disclosed that it utilized 45 trillion tokens across multiple modalities. However, the development process involved some use of distillation—the practice of training on outputs from other models. According to TechCrunch, Thinking Machines acknowledged that it used open-weight models, specifically Moonshot AI’s Kimi K2.5, to generate early post-training data. The company stated that it intends to move to fully self-contained post-training for future iterations. This transparency regarding training methodology stands in contrast to the often-opaque practices of larger labs, though the company remains guarded about its financial situation and the costs associated with its Nvidia-backed computing infrastructure.
The company’s focus on efficiency and customization is being tested against the scale of established competitors. While OpenAI and Anthropic have focused on building agentic, autonomous features atop general-purpose chatbots, Thinking Machines is betting that organizations will prioritize private, adaptable infrastructure. The startup currently employs approximately 200 people. As the industry debates the utility of frontier models versus specialized, open-source alternatives, the success of Inkling will likely hinge on whether enterprises find the burden of self-managed fine-tuning—which requires significant machine learning expertise—to be a viable trade-off for the control and cost-efficiency the company promises.