Saturday, November 21, 2026

INFORMOPEDIA

World news and deep-dive explainers — climate, economy, technology, science, culture, health

Breaking
Technology

Thinking Machines Releases Inkling 2 Mini, an Open-Weights AI Model That Runs on a Laptop

Thinking Machines Lab has released a capable AI model that fits on a laptop. Inkling 2 Mini, announced on Friday, November 20, is a 26.4-billion-parameter model with open weights that can run entirely offline on a consumer machine — and the lab says it beats every comparable open-weights model on tests of agentic coding.

The release is the second Inkling model in a month from the young lab founded by former OpenAI chief technology officer Mira Murati. Where Inkling 2 Small, released on November 5, is a 594-billion-parameter system that only runs on the lab’s own platform, Mini is built for the opposite end of the scale: a model small enough to download, own, and run at home.

A model that loops instead of growing

Inkling 2 Mini’s trick is architectural. It is a looped transformer: its 24 layers are divided into four blocks of six, and the two middle blocks run three times for every token, sharing the same weights on each pass. That gives 48 layer applications from 24 layers of weights, letting the model spend more computation on reasoning without adding parameters.

The model is a mixture-of-experts system with 26.4 billion parameters in total, of which 3.4 billion are active for any given token. Counting the extra work done by the loops, the lab puts the effective active count at 5.4 billion per token. Most of the weights were trained and shipped natively in NVFP4, a 4-bit format, with the last four layers in 8-bit MXFP8 — which is why the whole model fits in about 20 gigabytes of memory.

It is multimodal, accepting text and image inputs and producing text, and it supports a context window of up to one million tokens, according to its model card on Hugging Face. The weights are released under the Apache 2.0 licence, which permits commercial use, modification and redistribution — making Mini the first model in the Inkling 2 family anyone can actually download today.

Built a flight simulator on a plane

The lab’s demonstration is deliberately mundane in setting and striking in implication. Its researchers took Inkling 2 Mini on a flight, where it ran offline in airplane mode on a MacBook Pro with the M5 Pro chip and 48 gigabytes of memory, and built a playable flight simulator in a few minutes. No internet, no data centre, no per-token bill.

The model needs a machine with at least 32 gigabytes of memory: Apple silicon Macs with 32GB or more, NVIDIA’s DGX Spark desktop, or a GPU with 32GB or more of video memory. It runs through local inference tools including Ollama, vLLM, llama.cpp and SGLang, and the lab also offers it through its Tinker platform for inference and fine-tuning.

Leading its size class on coding

On the lab’s benchmarks, Inkling 2 Mini leads every open-weights model with under 100 billion parameters on agentic coding tests. It scores 48.2% on DeepSWE and 79.2% on SWE-bench Verified — against 46.1% for Google’s Gemma 4 31B and 77.8% for Alibaba’s Qwen3.8-27B, its closest rival. On BrowseComp, a test of deep web research, it reaches 79.0%.

The picture is more mixed elsewhere. It trails Qwen3.8-27B on some knowledge-work tests and on multi-turn conversation, scoring 50.0% on MultiChallenge, and its factual knowledge is limited by its size — the lab says looping improves reasoning but does not add facts. It is also less reliable than larger models when using tools, and the lab recommends running its agentic work in a sandbox and requiring confirmation before it deletes or overwrites files.

Why it matters

The release lands at the end of a crowded fortnight for AI models. OpenAI’s GPT-6.2 Astra and Sol arrived on November 17, and Google’s Gemini 4 Neon followed a day later — both closed models, available only through their makers’ services. Inkling 2 Mini points the other way: a model you can run yourself, offline, on hardware you already own.

That matters for privacy, for cost, and for control: a local model keeps documents and code on the machine, charges nothing per token, and can be fine-tuned without asking a provider. Stanford researchers who tested it early found it unusually efficient, scoring 0.703 on their intelligence-per-joule measure on a DGX Spark, against 0.430 for Qwen’s comparable model.

What to watch: whether independent researchers reproduce the lab’s benchmark numbers now that anyone can run the model, how quickly it appears in local AI tools, whether the bigger Inkling 2 Small’s weights follow — and what outside researchers find when they study the variable-loop model released alongside Mini to test whether looped models can still be monitored.

Avatar photo
Informopedia News Desk

Join the discussion

Your email address will not be published. Required fields are marked *