
Thinking Machines Lab has released Inkling 2 Mini, an open-weights artificial intelligence model designed to run on consumer laptops, saying its architecture delivers strong coding and reasoning performance while using less memory than larger systems.
The model has 26.4 billion total parameters and uses a looping transformer design that runs its middle layers three times for each token. The company says repeating these layers allows the model to carry out more computation internally, improving reasoning without requiring it to generate as many tokens.
Now loading...
On DeepSWE v1.1, a benchmark for agentic coding, Inkling 2 Mini scored 48.2%, ahead of other open-weights models with fewer than 100 billion parameters, according to the company. It also scored 77.9% on the hard subset of SWE-bench Pro v2. Thinking Machines Lab said the model produced its DeepSWE result at a theoretical maximum speed of 54.4 tokens per second, faster than similarly performing alternatives. Benchmark results can vary with evaluation methods and configurations.
The model is a mixture-of-experts system, with 5.4 billion weights activated per token, and has a memory footprint of about 20 gigabytes. Most weights were trained and deployed using four-bit precision, with the final four layers using eight-bit precision. The company said it runs on supported Mac computers with at least 32GB of memory and on NVIDIA’s DGX Spark desktop system.
Researchers also tested a separate variable-loop version, which can run with different numbers of passes. In tests without a written chain of thought, its estimated task time horizon rose from 16 seconds after one pass to 37 seconds after three. On a serial lookup task, accuracy continued to enhance with additional passes, including up to ten, although gains on many other tasks levelled off after three.
The company said the variable-loop model’s reasoning traces became shorter as the number of passes increased but remained readable in its initial analysis. Average reasoning length on a set of questions fell from 3,703 tokens at one pass to 1,661 tokens at five, while measures of sentence length and English word usage stayed similar. Researchers also found no clear increase in performance on tests of agentic sabotage as loop depth rose.
Those findings are preliminary, and the company said its experiments do not establish how models trained with reinforcement learning at higher loop counts would behave. In tests of whether the model could deliberately reshape its chain of thought to follow instructions, Inkling 2 Mini and the research model showed very low controllability, making it difficult to identify differences between loop settings. In agentic safety tests, a monitoring model detected almost all malicious side tasks when it could inspect both the system’s text and reasoning, but missed more when limited to tool activity. Researchers reported no clear pattern linking the miss rate to the number of loops.
Thinking Machines Lab said looping improved performance on long-context and multi-step reasoning tasks but did not add factual knowledge, which it said largely depends on a model’s total parameter count. The model performed strongly on search and tool-use benchmarks, including BrowseComp, DeepSearchQA and MCP Atlas, while trailing some alternatives on tasks involving complex instructions and multi-turn chat.
The company is releasing the model weights and making Inkling 2 Mini available for inference and customization through Tinker. It can also be run using several open-source inference tools, including Ollama, and the weights for the model and its variable-loop research variant are available on Hugging Face. The company said it hopes the release will help researchers investigate looped transformers and how their internal reasoning can be monitored.
