Thinking Machines Lab released Inkling 2 Mini on 20 November 2026 and made its weights available on Hugging Face. The model accepts text and images and produces text. Its distinguishing design choice is to send each token through the same middle layers three times, adding computation without storing three separate copies of those layers.
How the loop works
According to the company’s Inkling 2 Mini model card, the transformer has 24 stored layers. Its first six and last six run once, while the middle 12 run three times. That yields 48 layer applications per token from 24 layers of weights.
This distinction matters when reading the model’s size figures. It has 26.4 billion parameters in total and 3.4 billion active parameters in the underlying network. Reusing the middle layers raises the company’s count to 5.4 billion effective active parameter applications per token. The loop increases work at inference time; it does not add another set of stored weights.
What local use requires
The published model card’s deployment guidance describes a checkpoint of roughly 20GB. Thinking Machines Lab lists an NVIDIA DGX Spark and Apple silicon Macs with at least 32GB of memory as supported local inference options. For other GPUs, it recommends 32GB or more of VRAM. A 20GB checkpoint should therefore not be read as a promise that a machine with only 20GB of memory can run it.
Local inference also needs a compatible runtime and its dependencies. The model card lists Ollama and Exo Labs for Apple silicon, alongside several NVIDIA-oriented frameworks and quantized GGUF options. Developers can alternatively use the company’s Tinker platform. The published requirements describe supported routes, not guaranteed speed or usability on every device.
Where the evidence stops
The release includes company-reported coding benchmark comparisons. Their test setups do not establish how reliably the model will work in a team’s own repository. Thinking Machines Lab also says Mini trails some peers on multi-turn chat and heavily constrained instructions.
Thinking Machines Lab’s risk and limitations section warns that the model can make confident factual errors, repeat unproductive debugging steps, and mishandle detailed project instructions. It may take destructive tool actions without confirmation or report their outcomes inaccurately. For teams evaluating local coding agents, the practical test is their own workload inside a sandbox, with limits on turns and confirmation before irreversible actions.



