Thinking Machines Launches Inkling 2 Mini: A New Open Looped Transformer

Thinking Machines has unveiled Inkling 2 Mini, a new open looped transformer model. This model operates by running 12 of its 24 layers three times for each token, allowing it to reason internally for longer periods without generating additional text. This approach enables the model to arrive at answers using fewer tokens, according to Thinking Machines.
Inkling 2 Mini features a total of 26.4 billion parameters, with 5.4 billion active per token due to its looping mechanism. It is capable of processing both text and images, boasts a context window of 1 million tokens, and is licensed under Apache 2.0.
In performance tests, Inkling 2 Mini leads all open models under 100 billion parameters, achieving a score of 48.2 percent on the DeepSWE v1.1 coding test and 46.5 percent on the Vals Index v2.2. According to Artificial Analysis, it scores 20 points on its Intelligence Index, tying with Qwen3.8 27B while being approximately six times less expensive and four times faster.
The model runs locally on Macs with a minimum of 32 GB of memory and on Nvidia’s DGX Spark. Users can download the model weights and a research version featuring variable loops from Hugging Face. Additionally, it is available for testing in the Tinker Playground.
Source: The Decoder