Thinking Machines launches Inkling 2 Mini for laptop use
The open-weights AI model repeats internal processing layers, an approach the lab is also studying for its implications for monitoring.

Thinking Machines Lab has released Inkling 2 Mini, an open-weights AI model designed to run on laptops. The startup founded by Mira Murati says the model outperforms all open-weights competitors below 100bn parameters on coding benchmarks including DeepSWE.
The model repeatedly processes information through its middle layers, running them three times per token while its outer blocks run once. It has 26.4bn parameters, with 3.4bn active for each token, and occupies about 20GB of memory. Its weights are available through Hugging Face under an Apache 2.0 license. Mac users can run it through Ollama if their machines have at least 32GB of memory.
Thinking Machines reported scores of 48.2% on DeepSWE v1.1 and 77.9% on a difficult selection of SWE-bench Pro tasks. Independent testing firm Artificial Analysis assigned it an Intelligence Index score of 20, matching Qwen3.8 27B while using about five times fewer active parameters.
The lab acknowledged limitations: Qwen3.8-27B performs better in multi-turn conversations and on instructions containing numerous constraints. Inkling 2 Mini also falls behind larger models on difficult reasoning tests.
The looping approach raises a monitoring question because more reasoning can happen internally rather than appear in written output. Thinking Machines is also releasing a research version supporting one to five loops for outside study.
The lab said its preliminary tests found no meaningful decline in how well the written reasoning could be monitored at this scale. However, a monitor restricted to tool calls and outputs failed to detect a significant number of attacks.




