Thinking Machines Labs Inc. has released a new large language model that can automate artificial intelligence training tasks.

The algorithm, Inkling 2 Small, made its debut late Thursday. Around the same time, Business Insider reported that Thinking Machine is in the process of raising a $5 billion funding round led by Accel. The company, which was founded last year by former OpenAI Group PBC Chief Technology Officer Mira Murati, is expected to receive a $40 billion valuation.

Thinking Machines trained Inkling 2 Small on a cluster of GB300 NVL72 racks. Each system contains 72 Blackwell Ultra graphics cards and half as many central processing units. The company completed the initial pre-training phase in 68 hours and subsequently refined the model through a nine-day reinforcement learning workflow. There were also a number of intermediate training steps.

Thinking Machines honed Inkling 2 Small’s programming capabilities by giving it hundreds of thousands of long-horizon coding challenges. The task collection included a number of particularly complicated “autoresearch” exercises. They required the model to post-train neural networks, optimize LLM components called kernels and perform other AI research work.

Thinking Machines put Inkling 2 Small’s machine learning skills to the test by asking it to fine-tune Qwen3.5-4B. The goal was to improve the open-source LLM’s ability to make predictions such as weather forecasts. According to the company, Inking 2 Small doubled the model’s forecasting score in 4.5 hours.

The LLM is also adapt at automating business tasks. It outperformed Spark Open, Kimi K3 and several other open-source models on three popular knowledge work benchmarks. Additionally, it earned higher scores across several financial task datasets.

Inkling 2 Small includes a vision encoder that can analyze images. Moreover, it enables the LLM to visually verify the quality of the graphics that it generates. If a user asks the model to create an e-commerce store, it can not only check the website’s code for bugs but also visually inspect product pages to check that they render correctly.

The vision encoder accounts for 292 million of Inking 2 Small’s 594 billion parameters. About 17 billion of those parameters activate when a user enters a prompt. Inking 2 Small can adjust the amount of computing power that it spends on tasks, which enables developers to trade off output quality for increased cost-efficiency or vice versa.

Particularly savvy users can extend the model’s capabilities by post-training, or customizing it, using their own data. Thinking Machines says that Inkling 2 Small learns faster than its earlier models. That should translate into more cost-efficient post-training runs. 

Inkling 2 Small is available via Thinking Machines’ Tinker cloud service on launch. The company plans to release the model’s weights under an open-source license next week.

Image: Unsplash