i Informopedia.com
Artificial Intelligence Tier 1 · Public

How to Run Inkling 2 Mini on a Laptop: Requirements, Setup, and What It Can Do

2,297 words 11 min read Updated November 21, 2026

Inkling 2 Mini is the first AI model from Thinking Machines Lab that you can run yourself, on your own computer, with no internet connection and no bill per token. Released on Friday 20 November 2026, it is a 26.4-billion-parameter model whose weights are free to download under the Apache 2.0 licence — and it is small enough, at about 20 gigabytes, to fit on a well-equipped laptop.

That combination is unusual. Most models this capable live in data centres and are rented by the token. Mini is designed for the opposite: local, private, offline use, including as a coding agent that can work through a software project on your machine. This guide explains what the model is, what hardware it needs, how to run it, and where it falls short — so you can decide whether it belongs on your laptop.

A MacBook Pro laptop on a desk
A MacBook Pro on a desk. Thinking Machines Lab demonstrated Inkling 2 Mini running offline on a MacBook Pro. Photo: Pixel.la (CC0).

What is Inkling 2 Mini?

Inkling 2 Mini is a general-purpose, multimodal AI model from Thinking Machines Lab, the company founded by former OpenAI chief technology officer Mira Murati. It accepts text and image inputs and produces text, and it is built for agentic work: multi-step jobs where the model plans, uses tools, checks its own output, and keeps going without being prompted at every step.

Its defining feature is a looped transformer architecture. A standard transformer passes each token through its layers once. Mini has 24 layers, divided into four blocks of six, and the two middle blocks run three times for every token, reusing the same weights on each pass. The result is 48 layer applications from 24 layers of stored weights. The lab’s reasoning is that the middle layers carry the more abstract reasoning, so running them again buys more thinking per token without adding parameters.

The model is a mixture-of-experts system: each token is routed to 8 of 128 expert networks per layer. It has 26.4 billion parameters in total, but only 3.4 billion are active for any given token. Because the loops add extra work, the lab counts 5.4 billion effective active parameters per token. Most weights were trained and shipped in NVFP4, a 4-bit numeric format, with the last four layers in 8-bit MXFP8 — and because the weights are deployed in exactly the precision they were trained in, there is no quality lost to quantisation after the fact.

One more number matters: the context window is up to one million tokens, according to the model card on Hugging Face. That is the architecture’s capacity; in practice, long contexts need more memory and run more slowly, so treat it as headroom rather than a daily working size.

How it differs from Inkling 2 Small

Mini is the smaller sibling of Inkling 2 Small, released on 5 November. That model has 594 billion parameters, 17 billion active per token, and was trained for long-horizon autonomous work — tasks running for hours. It is also, for now, a hosted model: its weights were promised for the week after its announcement and have not yet been published, so the only way to use it is through the lab’s Tinker platform.

Mini reverses almost every part of that picture. It is roughly one-twentieth the size, it runs locally instead of in the cloud, and its weights are already public. What it keeps from its sibling is the emphasis on agentic coding and research: the lab says its strengths mirror Small’s, and on several benchmarks it comes surprisingly close to the larger model. On PMPP-Hard, a test of kernel optimisation, it nearly matches Small’s score. On factual knowledge it does not: looping makes a model reason better, the lab says, but it cannot store more facts — knowledge is mostly a function of total parameter count.

The practical difference is who each model is for. Small is a hosted service for teams with long, expensive jobs. Mini is a download for anyone with a capable machine and a reason to keep their AI local.

What hardware do you need?

The hard requirement is memory. The model checkpoint is about 20 gigabytes, and the lab’s model card lists three supported setups:

  • Apple silicon with 32GB of memory or more — MacBooks, Mac minis and Mac Studios. The lab’s own demo ran on a MacBook Pro with the M5 Pro chip and 48GB, in airplane mode, where it built a playable flight simulator in a few minutes.
  • One NVIDIA DGX Spark — the company’s 128GB desktop AI computer.
  • Other GPUs with 32GB of video memory or more.

A laptop with 16GB of memory will not run it comfortably; 32GB is the floor, and more is better, especially for long contexts. Storage is the easy part: 20GB for the checkpoint plus room for a runtime and a cache. There is no subscription and no account needed to download the weights — they are on Hugging Face, free, under the Apache 2.0 licence, which allows commercial use.

How to run Inkling 2 Mini

There are three main routes in, depending on what you want to do with it.

Ollama (the simplest route). Ollama is a free tool that downloads and serves open models with one command, and it supports Inkling 2 Mini on Apple silicon. Install Ollama, then pull the model and start chatting from your terminal. This is the fastest way to try it, and the best choice if you mainly want a local chatbot or a coding helper in your terminal.

vLLM or SGLang (the server route). If you want to serve the model to an application, or run agentic coding tools against it, vLLM and SGLang are the production-grade inference engines the model card lists, alongside Hugging Face’s own tooling, llama.cpp for quantised GGUF versions, TokenSpeed, NVIDIA Dynamo, and Exo Labs on Apple silicon. These need more setup than Ollama but give you control over batching, throughput and tool integration — the things that matter when the model is working as an agent rather than a chat window.

Tinker (the hosted route). If your machine does not meet the hardware bar, or you want to fine-tune the model, the lab’s own platform offers inference-only access through a playground, plus fine-tuning. This is the same platform that hosts Inkling 2 Small. It is the only route that involves an account and a bill; running the model locally costs nothing beyond electricity.

Whichever route you choose, start small: ask it to do a short, checkable coding task first, and watch how it behaves before handing it a whole project. The lab’s own advice is to set limits on output tokens and agent turns, because the model can get stuck re-checking its own work when debugging.

A laptop on a desk with books
A laptop on a desk. Inkling 2 Mini needs a machine with at least 32GB of memory to run locally. Photo: Freddie Marriage / Unsplash (CC0).

What it can do

Coding is the headline strength. Inkling 2 Mini leads every open-weights model with under 100 billion parameters on the lab’s agentic coding benchmarks: 48.2% on DeepSWE, 79.2% on SWE-bench Verified, and 77.9% on the hard tier of SWE-bench Pro. To put that in context, Google’s Gemma 4 31B scores 46.1% on SWE-bench Verified and Alibaba’s Qwen3.8-27B scores 77.8% — Mini beats both, and it is ahead of Qwen on DeepSWE by a wider margin still.

Research and tool use are the second strength. It leads similar-sized open models on BrowseComp (79.0%), a test of deep web research, on DeepSearchQA (85.3%), and on MCP Atlas (82.7%), a test of using external tools. That combination — code, search, tools — is what makes it useful as an agent rather than just a question-answerer.

Vision is solid rather than spectacular. It reads images and documents well — 92.8% on OmniDocBench, which tests understanding of documents — and it can look at what it has built and revise it, the way its bigger sibling does. On MMMU Pro, a broad test of visual reasoning, it scores 74.9%, just behind Qwen3.8-27B at 76.3%.

Efficiency is the quieter story. Stanford researchers who got early access measured intelligence per joule — how much energy a model needs per query — and found Mini scored 0.703 running on a DGX Spark, against 0.430 for Qwen’s comparable model and 0.159 for Gemma’s. The lab also demonstrated the point theatrically: twenty simulated villagers in a tiny pixel town called Thinkyville, all managed by one copy of Mini running on a single offline MacBook.

What it cannot do — the limits to know first

The lab is unusually candid about Mini’s weaknesses, and they are worth taking seriously before you rely on it.

First, it is not a great conversationalist. It scores 50.0% on MultiChallenge, a test of multi-turn conversation, and 33.4% on ComplexConstraints — behind Qwen3.8-27B on both. If you want a general chat assistant, a larger hosted model will serve you better.

Second, its factual knowledge is thin. Looping improves reasoning, not recall, and Mini’s score on AA-Omniscience, a factuality measure, sits in line with other models of its size. It is more likely than larger models to state wrong information confidently. It compensates by using search and tools well — which is why the research benchmarks look better than the closed-book ones — but it should not be trusted on facts it cannot check.

Third, agentic reliability is lower than with larger models. The lab warns that it can take destructive actions, such as deleting or overwriting files, without asking first, and that it may not report what it did accurately. Run it in a sandbox, require confirmation before irreversible actions, and keep backups. The same applies to its reasoning loops: in internal testing it sometimes got stuck re-checking its own work while debugging, burning tokens without progress.

Finally, local performance is still maturing. Support for looped models in local runtimes is new, so speed varies by runtime and hardware, and setting reasoning effort to its minimum does not fully switch reasoning off — the loops still run.

How Inkling 2 Mini compares

Against other open-weights models of similar size, Mini is at or near the top on most of the lab’s comparisons. It beats Google’s Gemma 4 31B by a wide margin almost everywhere — 79.2% against 46.1% on SWE-bench Verified, 85.3% against 61.7% on DeepSearchQA — and it edges Alibaba’s Qwen3.8-27B, its closest rival, on coding and research while trailing it on multi-turn chat and some knowledge-work tests. Against Meta’s Muse Glimmer it leads on most coding and tool-use benchmarks. The one test where Qwen clearly wins is AA-Briefcase, a test of professional knowledge work, where Qwen scores 1,391 to Mini’s 1,263.

Against closed models the picture reverses, as it does for every open model of this size. OpenAI’s GPT-6.2 Astra and Google’s Gemini 4 Neon, both released this week, lead on the hardest reasoning and on alignment, and come with mature APIs and ecosystems. The trade-off is the familiar one: open weights mean control, privacy and no per-token bill, at the cost of doing your own hosting, safety work and integration. For a closer look at another agent-style launch from the same period, see our guide to Perplexity Teammates, the AI agents for work.

Who should use it — and who should wait

Mini fits developers who want a local coding agent, researchers studying looped transformers, and anyone with private code or documents they would rather not send to a cloud service. It also fits tinkerers: the weights are free, the licence is permissive, and fine-tuning is supported through the lab’s platform and open tools.

Others should wait. If you need a general chat assistant, a hosted frontier model is better. If your machine has less than 32GB of memory, you cannot run it locally at all — use Tinker instead, or wait for smaller quantised versions from the community. And if you need reliable, unsupervised agentic work on important files, the lab’s own warning applies: sandbox it, supervise it, and do not let it act alone on anything you cannot undo.

The safety question: can we still watch it think?

There is a research question baked into this release, and the lab is open about it. The usual way to oversee an AI is to read its chain of thought — the reasoning it writes out as it works. Looping lets a model do more reasoning inside itself, without writing it down, which could in principle make it harder to monitor.

The lab tested this directly, releasing a variable-loop research model alongside Mini so outside researchers can repeat the work. Its findings are cautiously reassuring: as looping increases, the model does more without a written chain of thought — its no-CoT time horizon rose from 16 seconds at one pass to 37 seconds at three — but its reasoning stays legible, getting shorter and more direct rather than opaque. It also shows almost no ability to deliberately steer its chain of thought: it followed instructions about its own reasoning at most 0.1% of the time, against 0.4% for Inkling 2 Small and 1.3% for Qwen3.8-27B. In agentic tests it was unable to perform covert side tasks without detection. The lab’s conclusion is that at this scale, looping does not meaningfully reduce monitorability — but it is inviting the research community to check that claim rather than take its word for it.

What to watch next

The first thing to watch is whether independent researchers reproduce the lab’s benchmark numbers now that anyone can download and run the model. The second is how quickly it shows up in local AI tools — Ollama support is already there, and community quantised versions will follow. The third is whether Inkling 2 Small’s weights finally ship, completing the lab’s open-weights plan and giving developers the choice between a small local model and a large hosted one from the same family.

For now, Inkling 2 Mini is best understood as a practical milestone: the strongest small open model for coding you can run on your own laptop, with honest limits, a permissive licence, and a research question — how looped models think — that its makers want the world to help answer.

Related news coverage: our news report on the Inkling 2 Mini launch

More in Artificial Intelligence

Artificial Intelligence Tier 1 · Public

What Is Superhuman Go’s Personal Assistant? How the New Agent Works

Superhuman Go’s personal assistant is a named AI agent that works alongside you inside Superhuman’s productivity suite — watching your inbox, schedule and to-do list, flagging what needs attention, suggesting next steps,…

2,348 words 12 min read November 20, 2026
Artificial Intelligence Tier 1 · Public

What Is Google Genie 4? The World Model That Talks Back

Google Genie 4 is a world model: an AI that does not answer questions but builds a place you can walk through — and now, for the first time, a place you…

2,254 words 11 min read November 19, 2026
Artificial Intelligence Tier 1 · Public

What Is Gemini 4 Neon? Google’s Fast Frontier Model Explained

Gemini 4 Neon is Google’s answer to a practical question: what do you use when the everyday model is not quite enough and the flagship is more than you need? Announced on…

2,483 words 12 min read November 19, 2026