Low-poly laptop with a glowing looping ribbon of light above the keyboard, representing a small looped AI model running locally

Quick answer

Yes, mostly, for coding, agent and tool-use work. Inkling 2 Mini, released by Thinking Machines on 20 November 2026, ties Qwen 3.8 27B on the Artificial Analysis Intelligence Index (20.1 each) and beats it on most agentic coding and search benchmarks, while using far fewer active parameters. In WorldofAI’s hands-on tests it was usually faster, used fewer tokens and built better 3D and game demos. Qwen still wins on multi-turn chat, complex instructions and some knowledge tasks.

Key takeaways

  • Small enough for a laptop. About 20GB in size; Thinking Machines says it runs on a Mac with 32GB or more memory.
  • Open weights under Apache 2.0, so you can use, modify and deploy it commercially.
  • Looped design: its middle layers run three times per token, giving more reasoning without a bigger model.
  • Strong at agentic coding and search, for example 48.2% on DeepSWE v1.1 vs 42.2% for Qwen 3.8 27B, by Thinking Machines’ figures.
  • Weaker at long chats and many-rule instructions, where Qwen 3.8 27B scores higher.

🛠️ Practical guides

Ready to put this into practice? These step-by-step guides answer two common follow-up questions:

Most powerful AI models live in the cloud. Inkling 2 Mini is part of a new wave of open models that you can download and run on your own computer, and it is one of the strongest so far. This review covers what it is, how it compares with Qwen 3.8 27B, the obvious rival, and who it is worth trying for.

It is based on WorldofAI’s hands-on test video and Thinking Machines’ official announcement. Benchmark figures are as reported by Thinking Machines unless stated; we haven’t run our own benchmarks.

Watch: the source video

What this video covers: WorldofAI tests Thinking Machines’ Inkling 2 Mini, a 26-billion-parameter open-weight looped model, against Qwen 3.8 27B: benchmarks, local speeds on Macs and GPUs, quantised versions, and head-to-head builds including a zombie shooter, landing pages, a 360° product viewer, a soundboard, SVG art, a flight simulator and 3D scenes.

Video by WorldofAI (@intheworldofai). The video was sponsored by DecisionNode. We have no link with the sponsor.

What is Inkling 2 Mini?

Inkling 2 Mini is a small, open-weights AI model from Thinking Machines, built for coding, agent work, design and document tasks, and designed to run on a laptop.

SpecInkling 2 Mini
Released20 November 2026
Total parameters26.4 billion
Active per token3.4 billion per pass; 5.4 billion counting the loops
InputsText and images
Context window1 million tokens
Size on diskAbout 20GB (trained and shipped in 4-bit)
LicenceApache 2.0 (open weights)
Where to get itHugging Face, Ollama, Tinker and other tools

Because it is “Mixture-of-Experts”, only a small slice of the model works on each word, which is why it is fast for its size. For what the licence allows, see Can You Legally Modify an Open-Weight AI Model? Licences and Risks Explained.

What is a looped transformer, in plain English?

Instead of adding more layers to make a model think harder, a looped transformer runs some of its layers more than once. Inkling 2 Mini has 24 layers in four blocks of six. The first and last blocks run once; the two middle blocks run three times. So each word gets 48 layers’ worth of processing from 24 stored layers.

Think of it like re-reading the tricky middle of a problem a few times before answering. The result is better reasoning without a bigger download. The trade-off is that more “thinking” happens inside the model rather than in written-out reasoning. Thinking Machines says its tests found the model’s written reasoning got shorter with more loops but stayed readable, and it is releasing a research version so others can check.

💡 Tip: Looping helps reasoning, not memory for facts. A small model still knows fewer facts than a large one, which is why Inkling 2 Mini leans on web search and tools.

How does it compare with Qwen 3.8 27B on benchmarks?

They tie on the overall index, and Inkling 2 Mini leads on most coding and agent tasks. Selected figures from Thinking Machines’ table:

BenchmarkInkling 2 MiniQwen 3.8 27B
Artificial Analysis Intelligence Index20.120.1
DeepSWE v1.1 (agentic coding)48.2%42.2%
SWE-bench Verified79.2%77.8%
SWE-bench Pro v2 (Hard)77.9%73.8%
BrowseComp (web research)79.0%61.6%
MCP Atlas (tool use)82.7%77.3%
Terminal Bench v2.171.6%73.0%
AA-Omniscience non-hallucination rate63.4%71.4%

The pattern is clear: Inkling 2 Mini is better at doing things (coding, searching, using tools), and Qwen is a little better at knowing things and avoiding made-up facts. Read benchmark tables with care; each lab picks its own tests and setups.

What did the hands-on tests show?

WorldofAI ran both models through the same prompts. In the creator’s words, both are “incredibly close”, but Inkling 2 Mini came out ahead in most tests:

  • Zombie shooter game: Inkling 2 Mini built a playable game with rounds and areas; Qwen’s version was tilted and buggy.
  • 360° product viewer: Inkling 2 Mini produced a recognisable 3D shoe; Qwen’s 3D model was poor.
  • Landing page: both were average; the creator slightly preferred some of Qwen’s components, but Inkling was faster, used fewer tokens and was cheaper.
  • Soundboard: both did well; the creator liked Qwen’s sound slightly more.
  • SVG art and 3D scenes: an animated isometric room, a city skyline and a bee on a flower impressed the creator, who rated them above Claude Haiku 4.5 on 3D prompts.
  • Flight simulator: a working dashboard, though the plane wouldn’t take off in the creator’s test.
  • Practical coding (SQL, debugging): both passed; Inkling was usually faster and used fewer tokens.

These are one creator’s tests, not controlled benchmarks, but they match the official picture: strong at building, slightly weaker on polish and instructions.

How fast is it, and what do you need to run it?

HardwareReported speed
M5 Ultra Mac StudioAbout 210 tokens per second
M5 Max MacBook ProAbout 125 tokens per second
NVIDIA RTX 5090About 200 tokens per second

Those speeds come from the video and are well above reading speed. Thinking Machines and Ollama say you need a Mac with 32GB or more memory; it also runs on NVIDIA’s DGX Spark. Community members have made smaller quantised versions (around 5.4GB and 8.9GB) that reportedly keep most of the performance, though official weights are already 4-bit. For a full setup walkthrough, see How to Run Inkling 2 Mini on Your Mac With Ollama, and for the hardware question in general, How Much Memory Do You Need to Run an Open-Weight AI Model?.

⚠️ Warning: The video also points to an “uncensored” community version. It is unofficial, has had safety training removed and isn’t from Thinking Machines. We don’t recommend it for business use; see Uncensoring an AI Model: What Mark Kashef’s /unlock Skill Really Does, and the Risks for the risks.

Where does Inkling 2 Mini fall short?

  • Multi-turn chat and many-rule instructions: Thinking Machines itself says Qwen 3.8 27B scores higher here.
  • Factual knowledge: about what you’d expect for its size; use it with search for anything factual.
  • Hard reasoning: larger cloud models still do better on the toughest problems.
  • Hardware: 32GB of memory rules out many everyday laptops.

Who should use Inkling 2 Mini?

You are…Should you try it?
A developer wanting a local coding or agent modelYes, it is one of the best open options at this size
Working with private documents you can’t send to the cloudYes, if your computer can run it; see our offline guide
A beginner on a standard 8–16GB laptopNo, it won’t run well; use a cloud tool or a smaller model
Mainly chatting or writing long, rule-heavy contentMaybe; compare with Qwen 3.8 27B first

Before switching any real work, run your own tasks through it. Our guide on How to Test Whether a Cheaper AI Model Is Good Enough for Your Work shows a simple way.

Frequently asked questions

Is Inkling 2 Mini free?

Yes. The weights are free to download under the Apache 2.0 licence. You can also pay to use it through Thinking Machines’ Tinker platform.

Can Inkling 2 Mini run on a laptop?

Yes, on a Mac with 32GB or more memory, according to Thinking Machines and Ollama. It is about 20GB.

Is Inkling 2 Mini better than Qwen 3.8 27B?

For agentic coding, search and tool use, the reported benchmarks say yes. Qwen 3.8 27B is better at multi-turn chat, complex instructions and avoiding made-up facts.

What is a looped transformer?

A model that runs some of its layers several times for each word, giving more reasoning without making the model bigger.

Can I use Inkling 2 Mini commercially?

The Apache 2.0 licence allows commercial use, modification and redistribution, subject to its conditions.

Was the video sponsored?

Yes, by DecisionNode.

The bottom line

Inkling 2 Mini is a genuinely useful step for local AI: near-frontier coding and agent skills in a 20GB download that runs on a well-equipped laptop, with an open licence. It isn’t the best at everything, and it needs more memory than most everyday computers have. If you have the hardware and want private, offline AI for building and tool use, it is well worth testing.

author avatar
Garry Knight
I'm Garry Knight, the person behind Prodify Digital. I write about email list building, email marketing, SEO, AI search and the tools that connect them. My aim is to make online marketing easier to understand, so creators and small business owners can make informed decisions about building an audience and keeping people engaged. Here you'll find straightforward guides and product reviews that explain what something does, where it fits and which limitations matter. The focus is on clear explanations and useful next steps—not hype, shortcuts or promises of easy earnings.

Leave a Reply