Inkling 2 Mini, a local open model for repeated reasoning with 20GB
Inkling 2 Mini is an open-weights multimodal model from Thinking Machines Lab designed to increase reasoning computation through repeated middle layers while targeting local use with an approximately
Thinking Machines Lab released the open-weights model Inkling 2 Mini on November 20, 2026. The model’s key feature is a “looping transformer” architecture that passes the same intermediate layers through three times per token. Physically, it has 24 layers, but during execution it performs 48 layer operations, designed to carry out more computation even in short outputs. For developers who want to run open models on personal devices, this represents an important shift in the balance between size and inference performance.
Three passes through the intermediate layers
Inkling 2 Mini is an MoE model with a total of 26.4 billion parameters, of which 3.4 billion are activated per token. However, when repeated execution is taken into account, its effective activation is stated as 5.4 billion parameters. The six layers at the beginning and end run once each, while the 12 layers in the middle run three times, reusing their weights. In other words, it adds computation steps without greatly increasing the size of the model files. It accepts text and images as input and generates text, supporting a context length of up to 1 million tokens.
Conditions for local operation
The official model card states that the mixed-precision checkpoint is approximately 20GB. It recommends at least 32GB of memory on Apple Silicon devices and at least 32GB of VRAM in other environments. Therefore, it is more accurate to view it as a model aimed at local environments with sufficient unified memory or GPU memory, rather than a lightweight model that can run “on any ordinary laptop.” The weights are available from Hugging Face, while Tinker provides inference and customization.
Coding performance and remaining limitations
The developer emphasizes its strengths in agentic coding and tool use, reporting 48.2% on DeepSWE v1.1 and 46.5% on Vals Index v2.2. At the same time, it explains that the recurrent architecture does not itself increase the factual knowledge stored in the model. It may be weaker than larger models or some competing models in long, multi-turn conversations, instructions that must satisfy many conditions simultaneously, and questions with a high dependence on knowledge. When applying it to local agents, reliability should not be judged solely by the model’s shorter reasoning output; a setup that combines search and tool-call results with human review is necessary.