Thinking Machines unveils Inkling 2 Small LLM as $5B funding round emerges

Thinking Machines Labs Inc. has introduced a new large language model engineered to handle AI training automation. The model, called Inkling 2 Small, arrived Thursday evening, coinciding with reports that the company is pursuing a $5 billion funding round with Accel as lead investor. Founded last year by former OpenAI Chief Technology Officer Mira Murati, Thinking Machines is expected to reach a $40 billion valuation.

Training and capabilities

The company trained Inkling 2 Small using a cluster of GB300 NVL72 racks, each equipped with 72 Blackwell Ultra graphics cards and half that number of central processing units. Pre-training finished in 68 hours, followed by a nine-day reinforcement learning phase and additional intermediate training steps.

To develop Inkling 2 Small's coding abilities, Thinking Machines exposed it to hundreds of thousands of extended coding tasks, including complex autoresearch exercises that required the model to post-train neural networks, optimize LLM kernels and execute other AI research functions.

When tasked with fine-tuning Qwen3.5-4B to enhance weather forecasting predictions, Inkling 2 Small doubled the open-source model's forecasting score in 4.5 hours. The model also demonstrated strength in business automation, outperforming Spark Open, Kimi K3 and other open-source alternatives on three knowledge work benchmarks and scoring higher on multiple financial task datasets.

Vision and efficiency features

Inkling 2 Small incorporates a vision encoder capable of analyzing images and visually validating generated graphics. When building an e-commerce store, for instance, the model can check code for errors and visually confirm that product pages display correctly.

The vision encoder represents 292 million of the model's 594 billion total parameters. Around 17 billion parameters activate per user prompt. The model can modulate computational resources allocated to different tasks, allowing developers to optimize for either output quality or cost savings.

Advanced users can customize Inkling 2 Small through post-training with proprietary data. Thinking Machines reports that this version learns faster than earlier iterations, enabling more economical post-training workflows.

Availability

Inkling 2 Small launches today through Thinking Machines' Tinker cloud service. The company plans to release the model's weights under an open-source license the following week.