Free guide

Inkling 2 Mini: Run Thinking Machines' Looped AI Model on Your Mac

Thinking Machines' new open-weights model runs on a 32GB Mac with the wifi off. What it is, how the loop trick works, how to set it up, and the fine print.

By Beau Johnson·
The Short this guide goes with: Inkling 2 Mini: Run Thinking Machines' Looped AI Model on Your Mac

Quick answer

Inkling 2 Mini is a free, open-weights model (Apache 2.0) from Thinking Machines, released November 20, 2026. It has 26.4B parameters and runs its 12 middle layers three times for every token, so its 24 stored layers do 48 layers of work. It takes about 20GB and runs on an Apple silicon Mac with at least 32GB of memory through Ollama: ollama run inkling-2-mini.

The short version

  • Inkling 2 Mini is Thinking Machines' small open-weights model: 26.4B total parameters, 5.4B weight activations per token, Apache 2.0.
  • It's a looped transformer: the middle 12 of its 24 layers run three times per token. Thinking Machines says it leads every open model under 100B parameters on agentic coding tests such as DeepSWE v1.1 (48.2%).
  • It needs about 20GB and runs offline on a Mac with 32GB or more through Ollama. Their researchers built a playable flight simulator with it on a plane, in airplane mode.

How it actually works

A toy train that laps the middle loop

Picture a train set where every word is a train. A normal model sends each word through its stack of layers once. Inkling 2 Mini has 24 layers in four blocks of six: the first block runs once, the two middle blocks run three times, and the last block runs once. That's 48 layer applications per token from 24 stored layers.

Why loop at all

Thinking Machines says the middle layers carry the more abstract reasoning, so running them again lets the model apply more reasoning to each token and reach the right answer with fewer tokens. In their tests, looping the middle 12 layers beat looping the last 12 on 27 of 28 evaluations.

Small enough for a laptop

Most of its weights were trained and ship in 4-bit NVFP4 (the last four layers in 8-bit MXFP8), so there's no conversion step and no quantization loss. The whole model is about 20GB, which is why a 32GB Mac can run it.

Deeper, not more knowledgeable

Looping adds thinking, not memory. Thinking Machines says knowledge mostly depends on total parameter count, so its factual-recall scores are in line with other models its size. It makes up for that by using search and tools well.

How to set it up

  1. 1.Check your machine

    At launch, Thinking Machines supports any Mac (MacBook, Mac Mini) with at least 32GB of memory through Ollama, plus NVIDIA's DGX Spark. Ollama's release notes say it runs on Apple silicon only for now, with PC support coming in a later release. The download is about 20GB. Check your memory under Apple menu > About This Mac.

  2. 2.Install or update Ollama

    Inkling 2 Mini support arrived in Ollama 0.40.17. Download the latest Ollama from ollama.com, then check that your version is 0.40.17 or newer.

    ollama --version
  3. 3.Run it

    This downloads the model the first time, then opens a chat in your terminal. It accepts text and images and has a 1M token context window.

    ollama run inkling-2-mini
  4. 4.Use it as your coding agent

    Ollama can start a coding agent with Inkling 2 Mini as the model. The model page lists Claude Code, OpenCode, Hermes Agent and OpenClaw; Ollama's post also shows Pi. Planning to work offline? Install the agent, pull the model and sign in while you still have internet, then test it once with the wifi off before you rely on it.

    ollama launch claude --model inkling-2-mini
    ollama launch opencode --model inkling-2-mini
    ollama launch pi --model inkling-2-mini
  5. 5.No 32GB Mac? Try it in the browser

    Chat with it in the Tinker Playground, or use it on Tinker for inference and fine-tuning. The full weights are on Hugging Face (thinkingmachines/Inkling-2-Mini), and quantized GGUF versions are available through Unsloth and llama.cpp.

The fine print

  • Open weights under Apache 2.0, so the model is free to download, run locally and build on, including at work. Tinker, Thinking Machines' hosted service, is a separate paid product.
  • It's not the strongest small model for long multi-turn chat or instructions with many constraints: Thinking Machines reports 50.0% on MultiChallenge and 33.4% on ComplexConstraints, behind Qwen3.8-27B on both.
  • Looping doesn't add knowledge. For facts, give it search or tools instead of trusting its memory.
  • Benchmark numbers in this guide are Thinking Machines' own reported scores, with evaluation setups that vary by model.
  • Plan for about 20GB of disk space. On a 32GB Mac, close heavy apps while it runs.
  • Ollama runs it on Apple silicon Macs only in this release (0.40.17). PC support is coming in a later Ollama release; on NVIDIA hardware Thinking Machines points to DGX Spark, SGLang and vLLM.

FAQ

Is Inkling 2 Mini free?

Yes. The weights are open under the Apache 2.0 license on Hugging Face and Ollama. Running it locally costs nothing beyond your own hardware. Tinker, Thinking Machines' hosted service, is a separate product.

Can Inkling 2 Mini run on my laptop?

If it's a Mac with at least 32GB of memory, yes, through Ollama. The model takes about 20GB. Thinking Machines' own demo ran it offline on a MacBook Pro (M5 Pro, 48GB) in airplane mode.

What is a looped transformer?

A model that runs some of its layers more than once for each token. Inkling 2 Mini runs its middle 12 layers three times, so 24 stored layers do 48 layers of work, which gives it more reasoning per token without more parameters.

Is Inkling 2 Mini good for coding?

Thinking Machines says it leads every open-weights model under 100B total parameters on agentic coding benchmarks such as DeepSWE v1.1 (48.2%), and reports 79.2% on SWE-bench Verified (bash only). You can run it as the model behind Claude Code or OpenCode with ollama launch.

How is Inkling 2 Mini different from Inkling 2 Small?

Inkling 2 Small is far bigger (594B parameters) and scores higher on the Artificial Analysis Intelligence Index (31 vs 20.1). Inkling 2 Mini trades some of that for size: about 20GB, small enough for a 32GB Mac.

Sources

Facts checked 2026-11-21 against Thinking Machines' launch post (Nov 20), the Hugging Face model API, the Ollama model page, the Ollama v0.40.17 release notes and the @thinkymachines and @ollama launch posts.

Want help building with tools like this?

Join the Shipping Skool community free. Paid members get live bootcamp calls every week and weekly-updated courses.

Join Shipping Skool

More free guides