The most interesting AI release of the week did not come from OpenAI, Google, or Anthropic. It came from Thinking Machines Lab, the startup founded by former OpenAI chief technology officer Mira Murati, and it is a model designed to run not in a giant data centre, but on the laptop sitting on your desk.
The model is called Inkling 2 Mini, and it was released on 20 November 2026 with open weights under the permissive Apache 2.0 licence. That means anyone can download it, run it, study it, and build on it. It is the lab’s second open release in as many weeks, following Inkling 2 Small on 12 November, and it has already become the most popular new model on Hugging Face, the platform where AI developers share their work.
What makes Inkling 2 Mini unusual is not its size, but its shape. It is a looped transformer, a design that reuses the same parts of the model several times for each word it generates. The model has 24 layers, arranged in four blocks of six. The first and last blocks run once, as normal. The two middle blocks run three times each, with their weights shared across the repeated passes. The result is 48 layer applications per token, without adding any stored weights. In plain terms, the model thinks harder about each word without getting any bigger.
The numbers tell the story. Inkling 2 Mini has 26.4 billion parameters in total, with 3.4 billion active for each token, or 5.4 billion effective active parameters once the repeated passes are counted. The whole model fits in a checkpoint of about 20 gigabytes. The lab says it runs on a MacBook Pro with 48 gigabytes of memory, on a single Nvidia DGX Spark desktop AI computer, and on Apple silicon devices or graphics cards with 32 gigabytes or more of memory. Its researchers ran it offline on a MacBook Pro and had it build a playable flight simulator in a few minutes. That is the point. This is a capable model that does not need a cloud account, a data centre, or a monthly subscription.
It is also a capable model by the measures that matter to developers. It accepts text and images as input and produces text, with a context window of up to one million tokens, which means it can work with very long documents and codebases. In the lab’s own comparisons, it beats Qwen3.8-27B, one of the strongest open models of its size, on several demanding coding benchmarks, including DeepSWE, SWE-bench Pro v2 Hard, SWE-bench Verified, SWE-bench Multilingual, and Vibe Code Bench. Independent evaluator Artificial Analysis gave it a score of 20 on its Intelligence Index, and noted that it sits on the frontier of intelligence versus active parameters, meaning it gets more capability per unit of computing than its rivals. It costs 59 cents and takes nine minutes per benchmark task, against 3.53 dollars and 38 minutes for Qwen3.8-27B.
The training behind it is interesting, too. The lab started with 384 experts per layer, then pruned two-thirds of them to leave 128, and trained the model on one trillion tokens distilled from a larger internal Inkling model. Distillation, in which a smaller model learns from a bigger one, has become a standard way to pack capability into a small package, and Inkling 2 Mini is a good example of how far the technique has come.
For developers, the practical details are generous. The weights are available on Hugging Face, with ready-made versions for popular tools, including GGUF builds for running on ordinary computers, MLX support for Apple silicon, and day-one support in SGLang, vLLM, Ollama, and the Llama App. Fine-tuning is available through the lab’s own Tinker platform, as well as SkyRL, Miles, and Nvidia AutoModel. You can try it in the Tinker Playground without installing anything, or download it and run it entirely offline, which matters for people working with private code or sensitive data.
The release also says something about where AI is heading. For the past few years, the story has been about ever-bigger models running in ever-bigger data centres, accessible only through an internet connection and a monthly bill. Inkling 2 Mini points in the opposite direction. It suggests that clever architecture can do some of the work that raw size used to do, and that a model good enough for serious coding can live on a machine you own. That matters for privacy, because your code never has to leave your laptop. It matters for cost, because there is no per-token bill. And it matters for independence, because you are not renting intelligence from someone else’s server.
There are limits, of course. A 20-gigabyte model running on a laptop will not match the biggest frontier models on the hardest problems, and the lab is upfront that the extra passes through the looped layers mean more computation per token than the active parameter count alone suggests. It is also worth noting that the model’s strength is in agentic coding and tool use, the kind of step-by-step reasoning work that benefits from re-reading a problem. For other tasks, a conventional model of the same size may do just as well.
Even so, the reception has been enthusiastic. On its first day, Inkling 2 Mini became the number one trending model on Hugging Face, with more than 1,600 likes, while its larger sibling Inkling 2 Small sat at number 19. Developers are already quantising it, testing it, and building with it. One early report described it as dethroning Qwen’s model for local coding, which is exactly the contest Thinking Machines wants to win.
Mira Murati’s lab has been deliberately quiet since it launched, saying little and shipping carefully. With two open models in eight days, it is now speaking through its work. Inkling 2 Mini is not the biggest model of the year, and it is not trying to be. It is trying to be the model you can actually run yourself, on hardware you already own, and on that measure, it may be the most important release of the month.
