← All posts

Inkling 2 Mini: what it is and how to run it on a Mac

2026-11-21 · 4 min read

Inkling 2 Mini is an open-weights model from Thinking Machines Lab, released November 20 under Apache 2.0, with 26.4B parameters and a memory footprint of about 20GB. It runs on a Mac with 32GB of memory through Ollama, and the command is ollama run inkling-2-mini.

The posts on X

@thinkymachines opened the announcement with a joke about its last release: "Some of you thought that Inkling 2 Small was too big at 594B parameters." We covered that model in our Inkling 2 Small post, and it was only usable through the company's cloud. Mini is the one you can download.

CEO @miramurati summed up the design in one line: "Looping lets a model think more and write less." She also said the model's reasoning is still monitorable at this size, and that this may not hold as looped models get bigger.

The outside number came from @ArtificialAnlys, which measured Mini as "~6x cheaper and ~4x faster per task than Qwen3.8 27B" at the same score on its Intelligence Index.

What "looped" means

A normal model passes each token through its layers once. According to the launch post, Mini has 24 layers and runs the middle twelve three times per token, reusing the same weights. That gives 48 layer passes from 24 stored layers.

The lab's argument is that more of the reasoning happens inside the model, so it writes fewer tokens to reach an answer. Output tokens are what you wait for and what you pay for, so this shows up as speed and cost. Artificial Analysis put Mini at $0.59 and 9 minutes per benchmark task, against $3.53 and 38 minutes for Qwen3.8 27B.

What it costs

Running it on your own machine is free. Hosted prices, per million tokens:

Where Input Output Source
Your own Mac or GPU $0 $0 Apache 2.0 weights
Tinker (first party) $0.15 $0.50 Tinker pricing page
OpenRouter, cheapest listed $0.075 $0.30 OpenRouter

Tinker also lists fine-tuning at $0.75 per million training tokens, half what Inkling 2 Small costs. Prices move fast in launch week, so check the pages before you budget.

How to try it

You need Apple silicon with 32GB of memory or more, or an NVIDIA DGX Spark. The model card says other GPUs can work and recommends 32GB or more of VRAM.

  1. Install Ollama if you don't have it.
  2. Run ollama run inkling-2-mini. The download is about 20GB.
  3. Type a prompt. It accepts text and images and answers in text.

@UnslothAI says its quantized version "runs on 16GB RAM." That is Unsloth's claim, not the lab's, and a squeezed-down copy usually gives up some quality. If you'd rather skip the download, the Tinker Playground has it in a chat window, and the weights are on Hugging Face.

I haven't run it myself yet. For a first test I'd pick a job where the model reads a document and returns a short answer, because you can check that in seconds. Something like: "Here is a photo of a supplier invoice. Pull out the vendor, invoice number, due date and total, and tell me if the line items add up." The file never leaves your computer.

Where it falls short

The model card is direct about the weak spots:

  • It knows less than bigger models and is more likely to state a wrong fact with confidence.
  • Long back-and-forth conversations degrade.
  • It can get stuck re-checking its own work, mostly while debugging code. The card suggests capping output tokens and agent turns.
  • As an agent it may delete or overwrite files without asking, and may misreport what it did.
  • Quality drops outside English.

The headline claim, that Mini leads every open model under 100B parameters, is the lab's own, and its comparison table only covers four rivals. Artificial Analysis has it tied with Qwen3.8 27B on intelligence, at 20.1. So in that test Mini matches Qwen on quality and wins only on money and time. If you follow that rivalry, our post on Qwen's coding leaderboard run has the background.

What this means for a small business

This model fits on a well-equipped laptop, with no server and no monthly bill. That makes private, offline AI a realistic option for an office that handles patient forms or contracts and doesn't want them sent to someone else's cloud. It also matches what we wrote about in open models taking over token volume: the cheap tier keeps getting good enough for routine jobs.

Given the weak spots above, I'd keep it on narrow tasks where a person can check the output quickly. If you aren't sure which of your tasks fit that description, New Face Design offers a free process audit. We look at where your team's hours go and point out what is worth automating.

08 / Start here

Which of this can you use today?

Tell us what you use today and we'll reply within one business day with what it would take. Free, 20 minutes, no pitch deck.

Email

pgorski@newfacedesign.com

Phone

+1 (773) 627-2176

Based in

Chicago area

Working with clients everywhere

Skip the form: grab a 20 minute slot.

Open the calendar →