The Specialty News
AI

Thinking Machines' Inkling 2 Mini Fits in 20GB and Runs on a Laptop

Mira Murati's lab released the 26.4-billion-parameter model with open weights, and any Mac with 32GB of memory can run it offline.

By The Specialty News DeskEdited by 4 min read
Thinking Machines' Inkling 2 Mini Fits in 20GB and Runs on a Laptop
Photo: Thinking Machines Lab

On a MacBook Pro in airplane mode, with no internet at all, a new AI model wrote a working flight simulator. That model is Inkling 2 Mini, which Thinking Machines Lab released on Friday with open weights: 26.4 billion parameters packed into about 20GB, small enough for any Mac with 32GB of memory. For developers who want a capable coding assistant without sending a single line of code to the cloud, a laptop just became enough.

Three Laps Through the Middle Layers

Most language models pass each token through their layers once, top to bottom. Inkling 2 Mini goes around twice more. Its 24 layers sit in four blocks of six: the first and last blocks run once, and the 12 layers in the middle run three times for every token, so 24 stored layers do the work of 48.

The point is to do more thinking inside the model and less on the page. A reasoning model usually works through a problem by writing long chains of intermediate text, and every one of those tokens costs time and memory. Looping moves part of that work into the extra passes.

“Looping lets a model think more and write less.”— Mira Murati

The lab eased the loops in gradually during training through a learned gate, and each pass reads a short message left by the one before it. The harder trick was getting a model this capable down to laptop size in the first place.

From 594 Billion Parameters to 26.4 Billion

From 594 Billion Parameters to 26.4 Billion
Photo: Thinking Machines Lab

Earlier this month Thinking Machines released Inkling 2 Small, with 594 billion parameters. "Some of you thought that Inkling 2 Small was too big at 594B parameters," the lab wrote on X. Mini is the answer.

It started from the lab's main family of models. The team pruned two-thirds of the experts, keeping the 128 most used in each layer out of 384, then healed the model on 1 trillion tokens distilled from a larger internal Inkling, followed by 2.2 trillion tokens of reinforcement learning.

The size comes from precision as much as pruning. Most weights were trained and run natively in NVFP4, a 4-bit number format, with the last four layers in MXFP8. Because the model never has to be squeezed down after training, nothing is lost in conversion.

The whole thing runs on a MacBook or a Mac Mini with at least 32GB of RAM through Ollama, and on Nvidia's DGX Spark desktop. So what can it do once it's there?

A Flight Simulator Built in Airplane Mode

On the lab's own tests, Inkling 2 Mini scores 48.2% on DeepSWE, a benchmark of real software engineering tasks, against 42.2% for Alibaba's Qwen3.8-27B, and 79.2% on SWE-bench Verified. On the Artificial Analysis Intelligence Index it ties Qwen3.8-27B at 20.1, and Artificial Analysis measured each of its tasks at $0.59 and 9 minutes, against $3.53 and 38.2 minutes for Qwen.

Stanford researchers running it on a DGX Spark scored it 0.703 on their intelligence-per-watt measure, against 0.430 for Qwen3.6-35B-A3B. In another of the lab's demos, a single offline MacBook ran Thinkyville, a pixel-art town of 20 villagers, each one driven by the model.

The weights are on Hugging Face under the Apache 2.0 license, with a context window of up to 1 million tokens, and the model is live on Tinker, the lab's service for running and fine-tuning models. A hospital that can't send patient records to the cloud and a developer on a long flight now have the same option.

A top open coding model now fits in 20 gigabytes, and it works with the Wi-Fi off.

Same Index Score, 6x Cheaper

A visual summary of this story

The Brief

Stay curious

AI and technology: what changes and why it matters.
Your daily selection, in English or Spanish.

Free forever. Unsubscribe anytime.

Conversation

Start the conversation

No account needed. Comments are checked automatically — keep it civil.

More stories

Keep reading