Inkling 2 Small
Inkling 2 Small is Thinking Machines Lab's second-generation multimodal model for agentic coding, visual reasoning, and knowledge work. It accepts text and images, supports adjustable reasoning effort, and handles up to 1M tokens of context.
Model
Inkling 2 Small combines a sparse Mixture-of-Experts architecture with native vision support. It has 594B total parameters and activates 17B per token, routing each token to 8 of 384 experts across 48 decoder layers. Local and global attention layers support a context window of up to 1M tokens.
The model targets sustained agentic work, including software development, GPU kernel optimization, and machine-learning experiments. Thinking Machines trained it on long-horizon coding tasks across multiple agent harnesses and context-management strategies, with reinforcement-learning rewards that evaluate how it solves a problem as well as the result.
Vision capabilities include reading documents and charts, interpreting interfaces, and using visual feedback to refine websites, games, and 3D designs. For knowledge work, Inkling 2 Small can research information, use tools through MCP, and work with spreadsheets and financial documents. Thinking Machines reports 72.1% on DeepSWE v1.1 and 81.6% on MMMU Pro (Standard 10) in its release evaluations.