---
title: "Thinking Machines Lab’s Inkling 2 Mini is an open AI model small enough to run on a 32GB laptop"
url: https://badiyaa.com/inkling-2-mini-thinking-machines-laptop/
published: 2026-11-21T07:44:00+00:00
modified: 2026-11-21T07:44:00+00:00
section: Tech
publisher: Badiyaa
---

# Thinking Machines Lab’s Inkling 2 Mini is an open AI model small enough to run on a 32GB laptop

Thinking Machines Lab released Inkling 2 Mini, a 20GB open-weights AI model that runs offline on 32GB Macs and leads its size class on coding tests.

Thinking Machines Lab has released Inkling 2 Mini, an open-weights AI model whose whole checkpoint weighs about 20GB, small enough to run offline on a Mac with 32GB of memory. The company’s researchers say they ran it on a MacBook Pro in airplane mode during a flight and had it build a playable flight simulator in a few minutes.

The San Francisco startup, founded in February 2025 by former OpenAI chief technology officer Mira Murati according to [Wikipedia’s company profile](https://en.wikipedia.org/wiki/Thinking_Machines_Lab), published the model on November 20 under the Apache 2.0 license, which allows commercial use.

## The trick: the middle of the model runs three times

Inkling 2 Mini is what the lab calls a looped transformer. It has 24 layers split into four blocks of six. The first and last blocks run once, but the two middle blocks run three times for every token, so 24 stored layers do the work of 48. The idea builds on earlier research such as [the 2018 Universal Transformers paper on arXiv](https://arxiv.org/abs/1807.03819), which reused the same layers repeatedly.

The payoff, per [the lab’s launch post](https://thinkingmachines.ai/news/inkling-2-mini/), is more reasoning per token without a bigger download. The model has 26.4 billion total parameters and 3.4 billion active ones, which the looping stretches to 5.4 billion effective active parameters per token. The lab says looping the middle layers beat looping the last layers on 27 of 28 evaluations it ran.

It also trained most weights natively in NVIDIA’s 4-bit NVFP4 format, with the last four layers in 8-bit MXFP8, and ships them in that same precision. That means no extra compression step, and no quality lost to one.

## How it scores against rivals its size

Thinking Machines compares it with Google’s Gemma 4 31B, Alibaba’s Qwen3.8-27B and others under 100 billion parameters. On the DeepSWE v1.1 coding benchmark, Inkling 2 Mini scores 48.2% against 42.2% for Qwen3.8-27B. On SWE-bench Verified it reaches 79.2%.

It is not a clean sweep. The lab itself says the model trails Qwen3.8-27B in multi-turn chat and on instructions with many constraints, scoring 50.0% on MultiChallenge. Because looping does not add storage for facts, its knowledge scores sit in line with other models of the same size, and it leans on search tools to make up the gap.

## What you need to run it

[The official Hugging Face model card](https://huggingface.co/thinkingmachines/Inkling-2-Mini) lists three supported setups: a single NVIDIA DGX Spark, Apple silicon machines with at least 32GB of memory, or other GPUs with 32GB or more of video memory. Ollama, llama.cpp, vLLM, SGLang and Unsloth all support it, and the context window reaches 1 million tokens.

Here is the math for a laptop buyer. A 20GB model on a 32GB Mac leaves roughly 12GB for macOS, your apps and the model’s working memory, which grows with longer conversations. A 24GB machine is out; a 36GB or 48GB one gives breathing room. The scale shift is striking, too: the original Inkling released in July had 975 billion parameters, about 37 times as many as this one.

The model card also carries a blunt warning. When using tools, the model may delete or overwrite files without asking, so the lab recommends running agent jobs in a sandbox and requiring confirmation before anything destructive.

Badiyaa has covered the giant end of open models in [Alibaba’s 2.4-trillion-parameter Qwen3.8 release](https://badiyaa.com/alibaba-qwen3-8-2-4-trillion-open-model/) and the small end in [Pete Warden’s bet on local AI with Moonshine Voice](https://badiyaa.com/pete-warden-moonshine-voice-local-ai/).

Inkling 2 Mini is downloadable now, and the lab is also releasing a variable-loop research version so outside teams can study how looping changes what models do.

## Sources

- [Introducing Inkling 2 Mini](https://thinkingmachines.ai/news/inkling-2-mini/), Thinking Machines Lab, 2026

- [Inkling-2-Mini model card](https://huggingface.co/thinkingmachines/Inkling-2-Mini), Thinking Machines Lab on Hugging Face, 2026

- [Thinking Machines Lab](https://en.wikipedia.org/wiki/Thinking_Machines_Lab), Wikipedia, 2026

- [Universal Transformers (Dehghani et al.)](https://arxiv.org/abs/1807.03819), arXiv, 2018

*Image: Norbert Levajsics, CC0 via Wikimedia Commons*

*This article was produced with AI assistance and reviewed by our Editorial Desk. [How we work](https://badiyaa.com/editorial-policy/).*
