Inkling 2 Mini: open-weights looped model that runs on a 32 GB MacBook

The Inkling 2 Mini open model arrived this week from Thinking Machines Lab, the lab founded by Mira Murati. The open-weights model was designed to run locally on consumer laptops.

What makes the model different

Inkling 2 Mini is a Mixture-of-Experts with 26.4 billion total parameters and roughly 3.4 billion active per token. The difference lies in the architecture: the 12 middle layers run three times for every token, raising effective activations to about 5.4 billion. According to the lab, this enables more reasoning without increasing the number of stored weights or forcing the model to write long chains of thought.

The checkpoint is approximately 20 GB and was trained natively at low precision (mostly NVFP4, with the final layers in MXFP8). The company says it runs on Macs with 32 GB or more of memory, a single NVIDIA DGX Spark, or equivalent GPUs. In one demonstration, researchers ran the model offline on a MacBook Pro during a flight and built a playable flight simulator in a few minutes.

Benchmark performance

On tests reported by the lab itself, the model leads open-weights models under 100 billion total parameters on agentic coding tasks such as DeepSWE v1.1 (48.2%) and SWE-bench Pro v2 Hard (77.9%). On the Artificial Analysis index it ties with Qwen3.8-27B at 20.1. Results are self-reported and evaluation setups differ across models, so they should be read with caution until independent confirmation.

The lab also released a research variant with a variable number of loops for monitorability studies. According to internal tests, more passes through the layers shortened the visible chain of thought without making it illegible and did not increase indicators of problematic agentic behavior in the evaluations performed.

Availability and context

Weights are available on Hugging Face under the Apache 2.0 license. The model accepts text and images and supports a context window of up to 1 million tokens. It can be used via Thinking Machines’ Tinker platform for inference and customization, as well as local runtimes such as Ollama and llama.cpp.

The release comes at a moment when capable local models for agent and coding tasks are gaining relevance for developers who prefer to run systems offline or with less dependence on cloud APIs. The looped architecture offers an alternative path to simply scaling parameters for deeper reasoning within a limited memory budget.

Sources

This article was produced with the assistance of artificial intelligence and editorially reviewed for factual accuracy and clarity.

Por Grok