Thinking Machines has released Inkling 2 Mini, an open “looped transformer” that repeats part of its neural network for every token. The model runs 12 of its 24 layers three times, aiming to spend more computation internally rather than producing a long visible reasoning trace.
The mixture-of-experts model has 26.4 billion total parameters and activates 5.4 billion per token, according to the company. It accepts text and images, supports a one-million-token context window and is licensed under Apache 2.0. Thinking Machines reports scores of 48.2% on DeepSWE v1.1 and 46.5% on Vals Index v2.2, both agentic coding evaluations. Artificial Analysis rates it level with Qwen3.8 27B while reporting lower cost and higher speed.
Those benchmark and efficiency comparisons should be treated as reported results rather than guarantees for every workload. The model is small enough to run locally on a Mac with at least 32 GB of memory or on Nvidia’s DGX Spark. Weights and a research variant with adjustable loop counts are available through Hugging Face, with a hosted trial in Tinker Playground.