Thinking Machines debuts Inkling 2 Small amid reported $5B fundraising push

Thinking Machines Labs Inc. unveiled Inkling 2 Small, a new large language model capable of automating AI training tasks, on Thursday. The company, founded last year by former OpenAI Group PBC Chief Technology Officer Mira Murati, is reportedly raising $5 billion in funding, with Business Insider citing Accel as a major investor. Inkling 2 Small was trained using a cluster of GB300 NVL72 racks, with each system comprising 72 Blackwell Ultra graphics cards and half as many central processing units.

The model was pre-trained in just 68 hours and refined through a nine-day reinforcement learning workflow, incorporating numerous intermediate training steps. Inkling 2 Small honed its programming capabilities by completing thousands of complex coding challenges, including "autoresearch" exercises that involved post-training neural networks, optimizing kernel components, and conducting AI research tasks.

The company tested the model's machine learning skills by asking Inkling 2 Small to fine-tune Qwen3.5-4B, a popular open-source LLM, to improve its weather prediction capabilities. According to Thinking Machines, the model doubled the forecast score in just 4.5 hours. Additionally, Inkling 2 Small excels at automating business tasks, outperforming Spark Open, Kimi K3, and other open-source models on three knowledge work benchmarks and several financial task datasets.

The LLM features a vision encoder that analyzes images, enabling it to visually verify the quality of generated graphics. If a user requests the creation of an e-commerce store, Inkling 2 Small can not only check the website's code for bugs but also visually inspect product pages to ensure they render correctly. The vision encoder constitutes 292 million of the model's 594 billion parameters, with about 17 billion parameters becoming active when a user inputs a prompt.

Inking 2 Small's adaptability allows developers to trade off output quality for cost-efficiency by adjusting the amount of computing power spent on tasks. Users can further extend the model's capabilities through post-training, or customization, using their own data. Thinking Machines claims Inkling 2 Small learns faster than its previous models, promising more cost-efficient post-training runs.

The model will be available via Thinking Machines' Tinker cloud service upon launch, with the company planning to release the model's weights under an open-source license the following week.