egov.mn
Монголоор унших

Thinking Machines Lab releases Inkling 2 Mini, an open-weights looped model that runs on a laptop

Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, released Inkling 2 Mini, an open-weights looped transformer with a roughly 20GB footprint that it says leads open models under 100 billion parameters on agentic coding benchmarks.

This story was prepared automatically by software.

Official Thinking Machines Lab and Hugging Face graphic announcing Inkling 2 Mini on Hugging Face
Official Thinking Machines Lab and Hugging Face graphic announcing Inkling 2 Mini on Hugging Face

Thinking Machines Lab, the AI company founded by former OpenAI chief technology officer Mira Murati, on November 20 released Inkling 2 Mini, a compact but capable model small enough to run on a consumer laptop. The weights are available on Hugging Face under an Apache 2.0 license.

A looped transformer

Inkling 2 Mini is a looped transformer that runs the model's middle layers three times for every token. The company says this lets the model apply more reasoning internally, without writing it out, and reach correct answers with fewer tokens, especially on agentic coding tasks. It is a Mixture-of-Experts model with 26.4 billion total parameters and 3.4 billion active, and because most weights were trained and deployed natively in NVFP4, its total memory footprint is about 20GB.

Thinking Machines says Inkling 2 Mini leads every open-weights model under 100 billion total parameters on benchmarks such as DeepSWE v1.1 (48.2%) and Vals Index v2.2 (46.5%). Its researchers showed the model running offline on a MacBook Pro during a flight and building a playable flight simulator in a few minutes. The benchmark figures are company-reported and have not been independently reproduced.

Monitorability questions

Because looping lets a model reason longer without verbalizing its thinking, it could affect researchers' ability to oversee models by reading their chain of thought. Thinking Machines released a research variant with a variable number of loops to study this and reported that the chain of thought grew shorter with more loops but stayed legible, and that the model did not trend higher on tests of agentic sabotage.

The model is available on the company's Tinker platform and through open-source tools including SGLang, vLLM and Ollama.

Sources: Thinking Machines Lab, Hugging Face, Hugging Face model card

Share
Next storyChina-linked hackers' AI agents found a zero-day on their own and took over a server, Amazon says
Newsletter

Get the latest news delivered to your inbox.

Related articles