Thinking Machines has released Inkling 2 Mini, an open model that runs specific layers multiple times to extend its reasoning process without generating extra text. The system repeats 12 of its 24 layers three times per token, allowing it to find answers with fewer tokens while processing both text and images. It holds 26.4 billion parameters total, with 5.4 billion active per token due to this looping mechanism. The model operates under an Apache 2.0 licence and supports a context window of one million tokens. It runs locally on Macs with at least 32 GB of memory and on Nvidia DGX Spark hardware.
This architecture improves efficiency on agentic coding tasks compared to other open models under 100 billion parameters. Inkling 2 Mini scored 48.2 percent on DeepSWE v1.1 and 46.5 percent on the Vals Index v2.2. Independent analysis places it at 20 points on the Intelligence Index, matching Qwen3.8 27B performance at roughly six times lower cost and four times the speed. The weights are available for download from Hugging Face, including a research version with variable loop settings. Users can also test the model in the Tinker Playground.
- Runs on Macs with 32 GB RAM or Nvidia DGX Spark
- Apache 2.0 licence permits commercial use
- Research version allows variable loop configuration




