
Photo: AzureSaturn, 2025, via Wikimedia Commons (CC0)
For most of the past three years, the AI models worth talking about have lived in somebody else’s data center. You type, your words travel to a warehouse of graphics chips, and an answer comes back. On Friday, Nov. 20, 2026, Thinking Machines Lab, the San Francisco startup founded by former OpenAI chief technology officer Mira Murati, released a model built to work the other way around. Inkling 2 Mini is small enough to download, about 20GB, and run on a well-equipped laptop with no internet connection at all.
The short answer: Inkling 2 Mini is a free, open-weights model under the Apache 2.0 license, with 26.4 billion parameters, of which only 3.4 billion are used for each token, roughly a word or part of a word, that it writes. Its unusual trick is a "loop": twelve of its 24 layers run three times for every token, so the model does more thinking per word without needing a bigger file. Thinking Machines says it leads every open-weights model under 100 billion parameters on coding tests such as DeepSWE, and the independent benchmarking firm Artificial Analysis scores it level with Alibaba’s much heavier Qwen3.8 27B. To run it yourself you need an Apple silicon Mac with at least 32GB of memory, Nvidia’s DGX Spark desktop, or a graphics card with roughly 32GB of video memory. Everyone else can try it in the browser on the company’s Tinker Playground.
Thinking Machines announced the model on X at 12:10 p.m. Pacific time, and within the first hour more than 500 posts mentioned it, against a handful an hour all morning. Here is what was released, what the loop actually does, what hardware you need, how it scores, why the company spent half its announcement on a safety question, and what is still unknown. Prices were checked on November 20, 2026.
What Thinking Machines released
There are two downloads, both on Hugging Face. The main one, Inkling 2 Mini, is the finished model, tuned for coding, tool use and agent work, where the model takes a goal and works through many steps on its own. The second, TML-d24-varloop-research, has the same design but was trained to run its loop anywhere from one to five times. It exists so outside researchers can test how much the loop matters, and Thinking Machines says it is a research artifact rather than a product.
Both accept text and images and write text. Both can hold up to one million tokens of context, roughly the length of several long novels, in a single conversation. The license is Apache 2.0, one of the most permissive in software: businesses can use the model, change it and build it into paid products without asking anyone.
Support arrived on the same day across the tools people actually use to run models at home. Ollama, the popular Mac and PC app, added it in a pre-release of version 0.40.17. Hugging Face shipped support in its transformers library, llama.cpp and Apple’s MLX framework, and the model also runs on server software such as SGLang and vLLM. Hugging Face’s own launch write-up was enthusiastic but careful: "In our vibe tests, it’s at least comparable with Qwen 3.8 27B. More validation is needed."
What we know about how it works
A small model with a lot of experts
Inkling 2 Mini is a "mixture of experts" model. Instead of one big block of math that handles every word, each layer holds 128 smaller sub-networks, the experts, and a router picks 8 of them for each token. That is why the total size (26.4 billion parameters) and the working size (3.4 billion active parameters) are so different. You still need memory for the whole file, but each word costs only a fraction of the computing.
The model did not start that small. According to Thinking Machines, it began training with 384 experts per layer, and partway through training the team deleted two-thirds of them, keeping the 128 used most in each layer. It then trained the slimmed model on one trillion tokens of text generated by a larger internal Inkling model, so the surviving experts could take over the work of the ones that were removed. Fine-tuning and a round of reinforcement learning, 2.2 trillion tokens over 2,000 steps, came after that.
The loop

Diagram: Cosmia Nebula, 2024, via Wikimedia Commons (CC BY-SA 4.0)
This is the genuinely new part. A normal transformer, the design behind nearly every chatbot, passes each token once through a fixed stack of layers, from first to last. Inkling 2 Mini has 24 layers arranged as four blocks of six. The first block runs once, the two middle blocks run three times in a row with the same weights, and the last block runs once. So every token goes through 48 layer steps while the file stores only 24 layers.
The idea is that the middle of a model does most of the abstract reasoning, while the first and last layers mostly translate between words and the model’s internal representation. Running the middle again gives the model more room to work a problem through before it commits to the next word. In Thinking Machines’ tests, looping the middle 12 layers beat looping the last 12 on 27 of 28 evaluations. The company also compared looped and unlooped versions of the model before its experts were trimmed: the unlooped one was 7.5% cheaper to train but needed 140% more computing at run time to match the looped one on long agent transcripts.
Each pass is not a blind repeat. Every pass starts from the output of the first block, and the second and third passes also read a "message" from the previous pass through a learned gate. Early in training that gate is shut, so the model behaves as if it had no loop, and it opens gradually. Thinking Machines says this let it introduce looping without destabilizing training. Counting the repeats, the company describes the model as doing about 5.4 billion parameters’ worth of work per token.
Born in 4-bit
Most of the model’s weights were trained, not just shipped, in a compact 4-bit number format called NVFP4, with the last four layers in an 8-bit format called MXFP8. Usually a model is trained at higher precision and then squeezed down for home use, which costs some quality. Here the file you download is the same precision the model learned in, so there is no conversion loss. That is how a 26-billion-parameter model fits in about 20GB. Thinking Machines trained it on Nvidia’s GB300 systems and described the method in a Nov. 13 research post on end-to-end 4-bit training.
How well it does
On Thinking Machines’ own charts, Inkling 2 Mini resolves 48.2% of tasks on DeepSWE v1.1, a software-engineering benchmark, and scores 46.5% on the Vals Index v2.2. The company says both lead every open-weights model under 100 billion total parameters, and that it scores ahead of every comparable model on SWE-bench Pro v2 as well. Ollama’s model page lists 79.2% on SWE-Bench Verified, 82.7% on the MCP Atlas tool-use test and 76.5% on ScreenSpot Pro, a test of finding things on a screen.
The most useful outside number comes from Artificial Analysis, which runs the same battery of tests on every model it tracks. It put Inkling 2 Mini at 20 on its Intelligence Index, the same as Qwen3.8 27B at its extra-high effort setting and Nanbeige4.3-33B, while using about five and six times fewer active parameters respectively. The firm found that Inkling 2 Mini writes about 38% more words than Qwen3.8 27B to reach that score, roughly 152,000 output tokens per task against 110,000, but needs about 2.8 times less computing overall because each token is cheaper.
In dollars and minutes the gap is larger. Using Thinking Machines’ own API prices, Artificial Analysis measured $0.59 and 9.0 minutes per test task for Inkling 2 Mini, against $3.53 and 38.2 minutes for Qwen3.8 27B. For perspective, the same firm lists OpenAI’s GPT-6 Luna at its maximum setting at 28 on the index for $0.15 per task, so a small hosted model from a big lab can still be both smarter and cheaper. Inkling 2 Mini’s case is about what you can run yourself.
There is also an energy angle. Stanford researchers who measure "intelligence per joule," how much useful work a model does per unit of electricity, got early access. On a DGX Spark, Inkling 2 Mini scored 0.703, against 0.430 for Qwen3.6-35B-A3B, 0.159 for Gemma 4 26B A4B and 0.036 for Qwen3.5-27B. Thinking Machines notes those other models ran at different precisions, so the comparison is not perfectly like for like.
Where it falls short
Thinking Machines is unusually direct about the weak spots. Looping helps reasoning, but it does not let a model store more facts, which depend mostly on total size. On AA-Omniscience, a test of factual accuracy, Inkling 2 Mini is in line with other models its size, and closed-book trivia does not improve with extra loops. The company’s answer is that the model leans on search and tools, and it leads similar-sized open models on browsing tests such as BrowseComp and DeepSearchQA.
It is also not the best small model for long back-and-forth chats or for instructions with many constraints. It scores 50.0% on MultiChallenge and 33.4% on ComplexConstraints v1.1, behind Qwen3.8-27B on both, and it trails larger models on the hardest reasoning tests such as CritPt and HLE.
Looping helps reasoning, but it does not let a model store more facts.
Can you run it on your own computer?

Photo: Daniel Lu, 2026, via Wikimedia Commons (CC BY-SA 4.0)
Yes, if your computer has enough memory. The whole model has to fit in memory at once, and the file alone is about 20GB, so the official requirements are:
- Mac: any Apple silicon MacBook, Mac mini or Mac Studio with 32GB or more of memory, through Ollama, Exo or MLX.
- Nvidia DGX Spark: the company’s reference desktop, with 128GB of memory and native support for the 4-bit format.
- PC with a graphics card: other GPUs work, with 32GB or more of video memory recommended. Quantized GGUF files in 4-bit and 8-bit are available through Unsloth and llama.cpp.
- No hardware at all: chat with it in the Tinker Playground in a browser, or call it through Tinker’s API.
Will it run on a 16GB laptop?
Not as released. Thinking Machines and Ollama both list 32GB as the minimum for Macs. A 16GB machine cannot hold a 20GB file in memory with room left for the operating system and the conversation itself. Smaller community versions may appear, but none are official.
How fast is it?
Thinking Machines has not published a simple tokens-per-second figure for laptops. Its charts show a theoretical ceiling based on memory bandwidth, and the real speed will depend on your machine and settings. The practical sign is the company’s demo: researchers ran the model offline on a MacBook Pro during a flight and had it build a playable flight simulator in a few minutes. Another demo, a pixel-art village called Thinkyville, runs all 20 of its villagers on one offline MacBook.
What does it cost if you don’t run it yourself?
On Tinker, Thinking Machines’ developer platform, Inkling 2 Mini costs $0.15 per million input tokens, $0.015 per million cached input tokens and $0.50 per million output tokens. Fine-tuning costs $0.75 per million tokens with a 64,000-token context, or $1.50 with 256,000. Running the downloaded model at home costs nothing beyond your hardware and electricity.
Why would anyone run AI locally?
Privacy is the obvious reason: nothing you type leaves your machine. Cost is another for heavy users, since there is no per-token bill. And it works offline, on a plane or anywhere the connection is bad. The trade-off is that a model this size is not as capable as the largest hosted systems, and you are responsible for keeping it updated.
Why the loop raises a safety question
Roughly half of Thinking Machines’ announcement is not about speed or scores. It is about whether people can still keep an eye on a looped model.
One of the main ways labs check what a model is doing is to read its "chain of thought," the step-by-step reasoning it writes out before answering. That works partly because a normal transformer can only do so much computing per word before it has to write something down, so long reasoning ends up on the page where it can be read. A loop raises that limit. More of the thinking can happen inside the model, out of sight, which has led researchers to worry that looped models could be harder to supervise.
Thinking Machines tested this with the variable-loop research model. Answering with no written reasoning at all, the model’s "time horizon," the length of task it can finish half the time, rose from 16 seconds of human effort with one pass to 37 seconds with three, then leveled off. On a puzzle that requires following a chain of lookups in your head, it kept improving out to ten loops, twice what it was trained on. So the loop really does add hidden thinking.
The written reasoning, though, stayed readable. On science questions answered correctly at every setting, reasoning at five passes was less than half as long as at one pass, 1,661 tokens against 3,703 on average, but sentence length and the share of real dictionary words did not change. The model mostly stopped second-guessing itself. Asked to bend its chain of thought to follow an instruction, such as avoiding a word, Inkling 2 Mini managed it at most 0.1% of the time on mid-length reasoning, compared with 1.3% for Qwen3.8-27B. And in a test where a model tries to sneak a malicious side task past a monitor, there was no clear trend as loops were added.
Murati summed up the company’s position on X shortly after the launch:
Looping lets a model think more and write less. At this scale, its thought process remains monitorable; this may not stay the case.
That last clause matters. The tests cover a small model trained mainly at three passes. A much larger model trained to loop many more times might learn to hide more, and Thinking Machines says plainly that its experiments cannot rule that out. It has given early access to the interpretability lab Goodfire and is working with the National Deep Inference Fabric to make the looped models easier for academic researchers to study.
Where Thinking Machines fits
Thinking Machines Lab is barely 21 months old, and it has moved quickly from a secretive start to shipping a steady run of open models.
February 2025: Murati launches the company with about 30 researchers and engineers drawn from OpenAI, Meta and Mistral, including OpenAI co-founder John Schulman.
July 2025: A funding round led by Andreessen Horowitz raises about $2 billion at a $12 billion valuation, with Nvidia, AMD and Cisco among the investors.
Oct. 1, 2025: The first product, Tinker, a service for fine-tuning open-weights models, goes live.
March 10, 2026: Thinking Machines and Nvidia announce a long-term partnership to deploy a gigawatt of computing capacity.
July 15, 2026: Inkling, a 975-billion-parameter open-weights model, is released under Apache 2.0, followed on July 30 by the smaller Inkling-Small.
Nov. 5, 2026: Inkling 2 Small, 594 billion parameters with 17 billion active, opens the second generation and scores 31 on the Artificial Analysis index. Its weights reach Hugging Face on Nov. 12.
Nov. 20, 2026: Inkling 2 Mini, 26.4 billion parameters, the first of the family built for a laptop.
The company’s own launch post put it bluntly: some people thought Inkling 2 Small "was too big at 594B parameters." Mini is the answer. It also lands in a crowded field. Alibaba’s Qwen models, Google’s Gemma 4 and Nvidia’s Nemotron all compete for the same developers, and Thinking Machines’ own charts compare Mini most often with Qwen3.8-27B.
What’s still unclear
- Independent coding results. The headline DeepSWE and SWE-bench Pro figures are Thinking Machines’ own, and its charts note that evaluation setups vary by model. Outside leaderboards will take days to catch up.
- Real laptop speed. There is no official tokens-per-second figure for a 32GB MacBook, and the native 4-bit format is hardware-accelerated mainly on Nvidia’s newest chips.
- How the loop scales. The safety results cover a small model at mostly three passes. Whether bigger looped models stay as readable is an open question the company itself raises.
- Lower-memory versions. Whether useful builds for 16GB or 24GB machines appear, and how much quality they would lose, is not yet known.
The bottom line
Inkling 2 Mini is not the smartest AI model released this month, and Thinking Machines does not claim it is. What it offers is a new idea, looping a model’s middle layers to get more reasoning from a small file, packaged as a free model that a person with a 32GB Mac can download today and run on a plane. If the independent numbers hold up, it is among the most capable models you can keep entirely on your own machine. And because the company published both the model and a research version built to be pulled apart, it is also an open invitation to find out whether AI that thinks more and writes less is still AI we can watch.
Further reading

Build a Large Language Model (From Scratch)
Sebastian Raschka walks through coding a small GPT-style model step by step, the clearest way to see what layers, attention and training actually do.

AI Engineering: Building Applications with Foundation Models
Chip Huyen’s practical guide to building on models like Inkling 2 Mini, from choosing and evaluating a model to fine-tuning and serving it.

APC Back-UPS BE600M1 Battery Backup and Surge Protector
A small, popular battery backup that keeps a desktop AI computer and router running through short power cuts instead of crashing mid-job.

Logitech C920x HD Pro Webcam
A long-running 1080p webcam; Inkling 2 Mini reads images, so a desk camera is an easy way to feed it snapshots of whiteboards and documents.

Amazon Echo Dot
The cloud-based assistant most homes already have, a useful contrast to an AI model that runs entirely on your own machine.
