Thinking Machines releases Inkling 2 Mini
Thinking Machines has released Inkling 2 Mini, a fast, 26.4-billion parameter Mixture-of-Experts model featuring a looped transformer architecture designed for efficient local execution and strong reasoning capabilities. Running the middle layers multiple times allows the model to achieve high performance on agentic coding and reasoning benchmarks while maintaining a small memory footprint of approximately 20 gigabytes. Researchers found that looping increases internal computation without significantly reducing chain-of-thought monitorability in preliminary evaluations. Open article →