Thinking Machines Launches Inkling, an Open-Weight AI Model Challenging One-Size-Fits-All Approach

Thinking Machines Lab, the AI startup founded by former OpenAI CTO Mira Murati, has officially released Inkling, its first in-house AI model. Unlike the proprietary flagship models from major players like OpenAI, Anthropic, or Google, Inkling is an open-weight system, designed to allow external developers and companies to download and directly modify its parameters. This move signifies a strategic bet against the prevailing one-size-fits-all approach in the AI industry, emphasizing adaptability and customization for enterprise users.

Inkling is built as a mixture-of-experts (MoE) system, featuring 975 billion total parameters. However, for any given task, it efficiently utilizes only a fraction of these, approximately 41 billion, a common technique to maintain speed and cost-effectiveness in large models. The model was trained on an extensive dataset of 45 trillion tokens, encompassing text, image, audio, and video, enabling native reasoning across these modalities. While its foundational capabilities span multiple data types, Inkling’s initial outputs are focused on text, including code generation, styled artifacts, and structured data.

Key facts

Feature	Detail
Model Name	Inkling
Developer	Thinking Machines Lab
Model Type	Open-weight, Mixture-of-Experts (MoE)
Total Parameters	975 billion
Training Data	45 trillion tokens (text, image, audio, video)

A Year and a Half of Quiet Development

The launch of Inkling marks Thinking Machines Lab’s first public demonstration of its infrastructure development, a process that has largely taken place out of the public eye over the past 18 months. Earlier this year, the company offered a research preview of “interaction models,” showcasing AI designed for more dynamic and responsive communication, contrasting with traditional chatbot limitations. Inkling represents a significant validation of the startup’s core philosophy: that AI tailored to specific organizational needs will ultimately outperform generic, off-the-shelf models.

Thinking Machines emphasizes Inkling’s ability to provide calibrated answers, including flagging uncertainty rather than offering speculative responses. Users can also adjust the “thinking effort” parameter, allowing them to balance computational speed with desired accuracy. The company highlights that Inkling achieved comparable coding performance to Nvidia’s Nemotron 3 Ultra, an open-weight model, while using only a third of the tokens in one benchmark.

A Different Kind of AI

Thinking Machines explicitly states that Inkling is “not the strongest overall model available today, open or closed.” Instead, its focus is on well-rounded performance and flexibility. The company positions Inkling less as a finished product and more as a foundational tool that organizations can fine-tune through Tinker, its model-customization platform. This approach shifts some responsibility for safety and ethical deployment to the customers, who require significant machine learning expertise for effective customization.

This strategy diverges from the general-purpose chatbot development pursued by OpenAI (ChatGPT), Anthropic (Claude), and Google (Gemini), which prioritize broad applicability and agentic features. Thinking Machines argues that centrally trained, “set in stone” AI models often underperform compared to those shaped by an organization’s specific expertise. This perspective is gaining traction, with figures like Microsoft CEO Satya Nadella warning that enterprises using proprietary AI models effectively pay twice – through subscriptions and by contributing valuable business knowledge embedded in their prompts that can be absorbed into future model versions. Hugging Face CEO Clem Delangue has also predicted a shift where frontier models cater to experimentation, while most production AI work moves towards private or open-source alternatives, aligning with Thinking Machines’ vision.

Bridgewater Associates Project Illuminates Potential

A recent collaboration with Bridgewater Associates, the world’s largest hedge fund, provides a compelling example of Thinking Machines’ approach. Researchers from both entities enhanced an existing open-source model with Bridgewater’s financial expertise. This customized model reportedly scored 84.7% on financial reasoning tests, surpassing top proprietary AI models at approximately one-fourteenth of the running cost. While these results are from a joint evaluation, they underscore the potential for tailored AI solutions.

Thinking Machines also highlights its rapid development cycle, stating it reached this stage in about nine months, significantly faster than the five years for OpenAI or three years for Anthropic to bring their core technology to market and generate revenue.

Training Data and Costs

Regarding Inkling’s training data, Thinking Machines confirms a partial use of “distillation,” a practice of training models on outputs from competitors’ models. While Inkling was pre-trained from scratch, the company used other open-weight models, including Moonshot AI’s Kimi K2.5, to generate some early post-training data before large-scale reinforcement learning. The company intends for its next model to use fully self-contained post-training.

The financial aspects remain more guarded. Thinking Machines partnered with Nvidia in March to deploy a gigawatt of Vera Rubin computing capacity, training Inkling entirely on Nvidia’s GB300 NVL72 systems. However, details on how these substantial costs will be covered are not yet public, with revenue reportedly not being the primary focus. The company’s bet appears to be that its efficiency-driven approach will mean it doesn’t need to spend at the scale of its larger rivals, as the open weights mean customers aren’t obligated to pay Thinking Machines to run them. The revenue model is expected to stem from Tinker, the customization platform, through training, fine-tuning services, and a share of the hosting ecosystem. The company currently employs roughly 200 people.

For Indian enterprises and AI developers, Inkling’s open-weight nature offers a significant opportunity. It provides a foundational model that can be freely downloaded and adapted, potentially lowering barriers to entry for developing specialized AI applications. This approach could foster innovation in sectors requiring highly customized AI solutions, from finance to healthcare, allowing local businesses to build competitive advantages without being locked into proprietary ecosystems. The focus on efficiency and adaptability resonates with the needs of diverse Indian industries seeking cost-effective and tailored AI deployments.

Source: TechCrunch AI – https://techcrunch.com/2026/07/15/thinking-machines-amps-up-its-bet-against-one-size-fits-all-ai-with-its-first-open-model-inkling/