Beating AI News Flash: Thinking Machines' newly open-sourced Inkling 2 Small has 594 billion parameters and requires at least 500GB of VRAM to run. Many developers found it too large, so Thinking Machines released the smaller Inkling 2 Mini, with total parameters reduced to 26.4 billion and 3.4 billion different parameters activated per Token generation, finally able to run on an ordinary laptop.
Mini adopts the same recurrent Transformer architecture as GPT-6 Astra. The model has only 24 network layers, but the middle 12 layers run repeatedly 3 times. This means each Token generation goes through 48 layers of computation while only needing to store the weights of 24 layers. By reusing the same set of parameters, the model can perform more complex reasoning without continuing to expand in size.
Artificial Analysis composite score reaches 20.1 points, on par with Qwen3.8-27B, but the cost per individual evaluation is about one-sixth of the latter, and time taken is about one-quarter. Thinking Machines' published test results show that the DeepSWE v1.1 programming task pass rate reaches 48.2%, exceeding Qwen's 42.2%; however, in the AutomationBench-AA office automation task test, the score is 41.8%, still below Qwen's 48.2%.
The recurrent architecture also brings a problem: the model can complete more reasoning internally without writing out every step of thought, making it potentially harder for outsiders to inspect its reasoning process. Thinking Machines specifically trained a research model that can adjust the number of loops. Experiments found that the more loops, the more complex the tasks the model can complete without relying on textual reasoning, but no significant decline in the monitorability of the chain of thought has been found so far.
Inkling 2 Mini has released full weights under the Apache 2.0 license. The official version is about 20GB and can run offline on a Mac with 32GB of memory. Unsloth has also released a low-bit quantized version that can run on devices with 16GB of memory.
