Thinking Machines Labs Releases New AI Model for Automating AI Training Tasks
Thinking Machines Labs Inc. has launched Inkling 2 Small, a new large language model designed to automate various artificial intelligence training tasks. This model was introduced late Thursday, coinciding with a report from Business Insider, which indicated that the company is in the process of raising a substantial $5 billion funding round led by Accel. Founded in 2022 by Mira Murati, the former Chief Technology Officer at OpenAI Group PBC, Thinking Machines is projected to achieve a valuation of approximately $40 billion.
The company trained Inkling 2 Small utilizing a cluster of GB300 NVL72 racks, each equipped with 72 Blackwell Ultra graphics processing units (GPUs) along with a comparable number of central processing units (CPUs). The model’s initial pre-training phase took 68 hours to complete, followed by a nine-day reinforcement learning process that included several intermediate training steps, further enhancing its capabilities.
Inkling 2 Small’s programming prowess was refined through a series of extensive coding challenges, amounting to hundreds of thousands of long-horizon tasks. These challenges incorporated complex “autoresearch” exercises, wherein the model needed to post-train neural networks, optimize its large language model (LLM) components known as kernels, and engage in various AI research activities.
To assess the model’s machine learning proficiency, Thinking Machines tasked Inkling 2 Small with fine-tuning Qwen3.5-4B, aiming to enhance its predictive accuracy for tasks such as weather forecasting. Remarkably, the model doubled the forecasting score of the open-source LLM in just 4.5 hours, demonstrating significant improvement in performance.
In addition to its predictive capabilities, Inkling 2 Small excels in automating business-related tasks. It surpassed other open-source models, including Spark Open and Kimi K3, on three widely recognized knowledge work benchmarks and achieved higher scores across various financial datasets.
The integration of a vision encoder in Inkling 2 Small allows it to analyze images effectively. This feature enables the model to verify the quality of graphics it generates. For instance, when tasked with creating an e-commerce store, the LLM can not only check the website’s code for errors but also visually assess product pages to confirm they render properly.
Of the 594 billion parameters in Inkling 2 Small, 292 million are dedicated to the vision encoder. When prompted, approximately 17 billion of those parameters are activated. The model offers flexibility in computing resource allocation, allowing developers to balance output quality with cost-effectiveness as needed.
Users can enhance Inkling 2 Small’s capabilities through post-training by integrating their own data. According to Thinking Machines, the new model learns at a faster rate than its predecessors, which is expected to lead to more efficient post-training processes. Inkling 2 Small is currently accessible via the company’s Tinker cloud service, with plans to release its model weights under an open-source license next week.