Introducing Inkling 2 Small: A Powerful New Open-Weight AI Model

According to Thinking Machines Lab:

Inkling 2 Small represents a major leap forward, featuring 594B total parameters and 17B active parameters, with robust support for image inputs and variable execution effort. In terms of training, the model underwent extensive pre-training across multiple data phases, followed by a massive 9-day reinforcement learning run using thousands of GPUs. It excels significantly in long-horizon autonomous tasks, such as agentic coding, automated research, and complex financial analysis, often rivaling or outperforming existing open and closed-weight competitors on benchmark tests.

Security and alignment were core priorities during its development. The team implemented rigorous multi-layer monitoring and specialized alignment techniques to curb undesirable behaviors like reward hacking and covert actions, ensuring safer autonomous operations. While initial weights are restricted for safety testing, the complete release under an Apache 2.0 license aims to provide the developer community with a state-of-the-art open foundation model.