Thinking Machines’ agentic ambitions take a more compact form

An open-weights 26.4 billion-parameter model combines local execution, tool use, and the study of reasoning monitoring.

On an Apple silicon Mac with at least 32 GB of memory, Thinking Machines’ new model can be run locally. Released under the name Inkling 2 Mini, it takes up about 20 GB for its weights and accepts both text and images, with text responses. Its distinctive feature lies in its architecture: part of the layers is traversed multiple times before each token is produced.

The weights are downloadable on Hugging Face under an Apache 2.0 license. Ollama support still goes through its preview version on Mac. The maker also says it can run on DGX Spark, while Tinker offers inference, fine-tuning, and conversational access in its Playground. Twenty-four layers, forty-eight applications

The architecture includes 24 layers: the first six and the last six are run once, while the 12 in the middle are run three times. Since the repeated passes reuse the same weights, the model performs 48 layer applications per token without storing 48 different layers. Internal computation increases, without a corresponding increase in the number of parameters kept in memory.

The mixture-of-experts model has 26.4 billion parameters in total, of which 3.4 billion are active. Accounting for the repeated passes, the provider puts the number of weight activations at 5.4 billion per token. These two figures describe different things: the parameters engaged and the extra work tied to the loops.

The size was also reduced by removing two-thirds of the experts during training, then retraining the remaining experts to compensate for that removal. Low numerical precision also helps reduce the memory footprint: the weights are mostly trained and deployed in NVFP4, with MXFP8 for the last four layers. The model card announces a maximum context of one million tokens. The ranking changes depending on coding tasks

In the evaluations published by the provider, the model solves 48.2% of DeepSWE v1.1 tasks, compared with 42.2% for Qwen3.8-27B. On Terminal-Bench v2.1, the order is reversed: 71.6% versus 73%. The table also reports 20.1 on the Artificial Analysis Intelligence Index, tied with that competitor. The results depend on the configurations and tools used; the launch speed charts show a theoretical ceiling, without measuring throughput on each compatible computer.

The Artificial Analysis comparison also tempers the promise of reasoning with less text. At equal score, the model produces about 38% more tokens than Qwen3.8-27B. The organization nevertheless estimates a roughly 2.8x lower compute requirement with its indicator based on active parameters multiplied by tokens. A shortened chain of thought through loops therefore does not mean a response that is systematically more concise than that of another model. Is shorter reasoning still monitorable?

The internal passes raise a question: if more work happens before text is emitted, is the written chain of thought still useful for understanding the model’s behavior? A research variant with a variable number of loops is released to explore this point, while the user-facing model keeps three fixed passes.

On a subset of GPQA Diamond questions solved across all tested numbers of passes, the researchers observe a drop from 3,703 to 1,661 reasoning tokens on average between one and five passes. Sentence length and the proportion of words found in an English dictionary remain close. The analysis describes more direct reasoning, with fewer backtracks.

Tests in LinuxArena do not show any clear trend toward better concealment of malicious actions as the number of loops increases. For these tests, the study also uses a version of Mini retrained to accept tasks that the released model often refuses. The lab presents its observations as preliminary: the variable-loop variant was trained with one to five passes during part of its learning, but its reinforcement training used only three. The experiments therefore do not determine what a model optimized for more passes would learn. As a complement, a Goodfire study, conducted with early access to the weights, directly examines how information flows between the layers and the loops.