What changed
Thinking Machines Labs Inc. has released Inkling 2 Small, a large language model that can automate artificial intelligence training tasks. The model was announced alongside reports of a $5 billion funding round led by Accel, which is expected to result in a $40 billion valuation for the company.
What this means in practice
Inkling 2 Small was trained on a cluster of GB300 NVL72 racks, each containing 72 Blackwell Ultra graphics cards and 36 central processing units. The company completed the initial pre-training phase in 68 hours and refined the model through a nine-day reinforcement learning workflow. The model was also trained on hundreds of thousands of long-horizon coding challenges, including complex 'autoresearch' exercises that required it to post-train neural networks, optimize LLM components called kernels, and perform other AI research tasks.
The model was tested by fine-tuning Qwen3.5-4B, improving its ability to make predictions such as weather forecasts. According to the company, Inkling 2 Small doubled the model’s forecasting score in 4.5 hours. It also outperformed Spark Open, Kimi K3, and several other open-source models on three popular knowledge work benchmarks and earned higher scores across several financial task datasets.
Limitations and availability
Inkling 2 Small includes a vision encoder that can analyze images and visually verify the quality of generated graphics. The vision encoder accounts for 292 million of the model’s 594 billion parameters, with about 17 billion activating when a user enters a prompt. The model can adjust computing power usage to trade off output quality for cost-efficiency.
Particularly savvy users can extend the model’s capabilities by post-training or customizing it with their own data. Thinking Machines claims Inkling 2 Small learns faster than its earlier models, leading to more cost-efficient post-training runs. The model is available via Thinking Machines’ Tinker cloud service and will be released under an open-source license next week.
