Thinking Machines Lab, a new AI startup founded by Mira Murati, former chief technology officer of OpenAI, open sourced its new model Inkling 2 Mini on November 20, U.S. time. This model only requires about 20GB of memory and can be executed offline on a consumer-grade laptop. The researcher turned on airplane mode on the plane and used a MacBook Pro to write a playable flight simulator in a few minutes.

according to Thinking Machines official publicationIn the open source weight model with a total parameter of less than 100 billion, Inkling 2 Mini scored 48.2% in the DeepSWE v1.1 test and 46.5% in the comprehensive test Vals Index v2.2, both ranking first in its class.

The total parameters are 26.4 billion, only 3.4 billion are moved each time, and the memory is about 20GB.

Open source weights refer to the parameter files of the company's public model, which anyone can download to their own computer to execute or modify. according to Hugging Face model page, Inkling 2 Mini is licensed by Apache 2.0 and can be used commercially.

It is a "mixed expert" architecture: there are many groups of subnetworks in the model, each with its own expertise, and only a few of them are used at a time. The total parameters of Inkling 2 Mini are 26.4 billion, but only 3.4 billion are actually used for each word.

Most of the weights of the model use 4-bit NVFP4 format from training to deployment, and the last four layers use 8-bit MXFP8. There is no need to compress and convert during deployment, so there will be no points loss. Officials said that it will support Macs with 32GB or more of memory to run through Ollama when it is launched, and Huida's DGX Spark desktop AI computer can also run.

The "loop" design in which each word in the middle layer runs three times

The most special thing about Inkling 2 Mini is the loop architecture. The model has a total of 24 layers and is divided into 4 sections. Each time a word is processed, the first and last two sections are run once each, and the middle two sections are repeated three times. This is equivalent to using 24 layers of parameters to perform 48 layers of operations.

The official statement is that the middle layer is responsible for more abstract reasoning. Running it several times will allow the model to think more deeply and use fewer words to get the correct answer. This is especially helpful for AI agent programming tasks that require reasoning quality. In the official experiment, the loop was placed in the middle 12 layers, and 27 out of 28 evaluations showed more progress than placing the loop in the last 12 layers.

The cost is the amount of computation. Officially comparing the loop-free version of the same scale, the training cost of the loop-free version is 7.5% less, but to achieve the same performance, inference requires 140% more calculations.

I am ahead in writing programs, but I am not getting better at memorizing knowledge.

The official compares Gemma 4 31B, Qwen3.8-27B and other equivalent open source models. In the DeepSWE v1.1 programming test, Inkling 2 Mini scored 48.2%, and Qwen3.8-27B scored 42.2%;Hugging Face BlogIn summary, compared with these 4 models, it leads in 32 out of 48 tests.

Loops can enhance reasoning, but they do not make the model remember more knowledge. The scores for factual questions and answers are similar to those of models of the same size. Officials said that Inkling 2 Mini switched to using search and tools to supplement the results, and it performed ahead in tests such as BrowseComp that require online search for information.

The official also lists weaknesses: multiple rounds of dialogue and many conditional instructions, the score lags behind Qwen3.8-27B; in difficult reasoning tests such as HLE-Diamond, it is not as good as larger models.

If you think more but don’t write it out, will it be more difficult to supervise?

One of the main ways to supervise AI now is to read the reasoning process (thinking chain) written by the model. The loop allows the model to calculate a few more steps internally before speaking. The outside world is worried that some reasoning may not be written at all, making supervision more difficult.

Thinking Machines also open sourced a research version model with adjustable loop times for this purpose. The test shows that without writing inferences and answering directly, the length of the task that the model can complete increases from 16 seconds of human work when it is run once to 37 seconds when it is run three times, and then it is roughly the same.

When running five times, the average number of inference words of the model is 1,661 words (tokens), which is less than half of the 3,703 words when running once. However, the sentence length and the proportion of words in the English dictionary have not changed, and it is still understandable. In the LinuxArena test that simulated malicious tasks, the proportion of secretly completing malicious tasks without being caught did not show an obvious trend with the number of loops.

The official emphasized that this is only a preliminary result. The research version of the model's reinforcement learning only uses three loop trainings. A larger model with more loops may be able to compress reasoning even more. This is one of the reasons why they are open source, hoping that the research community will work together to study the supervisability of the loop model.

Inkling 2 Mini has been launched on Tinker, the fine-tuning platform of Thinking Machines. It can be inferred and customized, with full emphasis on Hugging Face. Common open source tools such as SGLang, vLLM, and Ollama are also supported.

this article The former chief technical officer of OpenAI created a new open source small model: the laptop runs offline and a flight simulator is written on the airplane. first appeared in Chain News ABMedia。