wasl.tech

Inkling 2 Mini: Thinking Machines’ Laptop-Sized AI Model Beats Bigger Rivals

Somewhere over the Atlantic, at 35,000 feet, a laptop sat in airplane mode. No Wi-Fi, no cloud, no data centre humming in the background. And yet the machine on the tray table was building a playable flight simulator — writing the code, testing it, fixing its own mistakes — in a matter of minutes.

The model doing the work was Inkling 2 Mini, and it had never left the laptop. Thinking Machines Lab released it on November 20, 2026. It is not the biggest model of the year — but it does something clever instead: it thinks the same thought three times.

That trick, called looping, lets a laptop-sized model outscore rivals many times its size on serious coding work — and its weights are open, so anyone can download it and run it offline.

What Is Inkling 2 Mini?

Inkling 2 Mini is a new open-weight AI model from Thinking Machines Lab, the startup founded by former OpenAI chief technology officer Mira Murati. It is the smaller sibling of Inkling 2 Small, a 594-billion-parameter model that remains closed for now.

Mini is a mixture-of-experts model with 26.4 billion total parameters, of which only 3.4 billion are active for any given token. Each token is routed to 8 of 128 experts, so most of the model stays asleep while a small slice does the work.

The weights are released under the Apache 2.0 licence on Hugging Face, which allows commercial use and fine-tuning without a custom agreement. It is also available on Tinker, the lab’s own platform, for hosted inference and fine-tuning, with chat access in the Tinker Playground.

Day-one serving support covers SGLang, vLLM, Ollama, Unsloth, NVIDIA Dynamo, TokenSpeed, and the Llama App. According to Thinking Machines Lab, the model is designed first and foremost for long-horizon agentic work — the multi-step coding and tool-using jobs that usually need a frontier model in the cloud.

How Inkling 2 Mini Loops Its Way to Smarter Answers

The headline idea behind Inkling 2 Mini is looping. The model has 24 layers, divided into four blocks of six. The first and last blocks run once per token, but the two middle blocks run three times, reusing the same weights on every pass.

That adds up to 48 layer applications from just 24 stored layers. The model gets more computation per token without carrying more weights around, which is exactly what a laptop-sized model needs.

The design builds on earlier research into looped transformers: the middle layers carry the abstract reasoning, while the first and last layers mostly translate between words and internal representations. Looping the middle is where the extra thinking pays off.

There is a second benefit, and it matters for safety. Looping lets the model reason longer without writing its reasoning down. Instead of producing a long chain of thought, it does more of the work silently, inside its own layers.

That raises a safety question: if the model thinks more without showing its work, can we still monitor it? The lab tested a variable-loop research variant to find out.

The results were reassuring. With three loops, its no-chain-of-thought time horizon rose from 16 to 37 seconds, and its reasoning stayed legible even as it grew shorter. The model showed no trend towards higher agentic sabotage as loops were added.

Inkling 2 Mini Benchmarks: Small Model, Big Scores

On the numbers Thinking Machines Lab reports, Inkling 2 Mini leads every open-weight model under 100 billion total parameters on several agentic coding tests. The headline score is 48.2% on DeepSWE v1.1, a demanding benchmark for real software engineering work.

It also scores 46.5% on the Vals Index v2.2 and 20.1 on the Artificial Analysis Intelligence Index, tying Qwen3.8-27B for the top score among models of its size. On SWE-bench Verified it reaches 79.2%, and on the harder SWE-bench Pro v2 it manages 77.9%.

The coding results keep coming. Inkling 2 Mini scores 73.7% on SWE-bench Multilingual and 71.6% on Terminal-Bench v2.1, both ahead of its closest open-weight rivals. On Vibe Code Bench it reaches 77.9%, well clear of Qwen3.8-27B at 64.9%.

It is not only a coder. The model leads its size class on BrowseComp at 79.0% and on MCP Atlas at 82.7%, two tests of search and tool use.

Artificial Analysis adds an efficiency angle. Inkling 2 Mini costs $0.59 and takes 9.0 minutes per Intelligence Index task, against $3.53 and 38.2 minutes for Qwen3.8-27B — roughly six times cheaper and four times faster for the same work.

All of these figures are vendor-reported, so independent testing will matter. But the gap over the previous open-weight champion, Qwen4-Max, is large enough to take seriously.

Running Inkling 2 Mini on a Laptop

The most surprising number in the release is not a benchmark at all. It is 20 gigabytes — the total memory footprint of the model. Most of its weights were trained and deployed natively in NVFP4 precision, with only the last four layers in MXFP8.

That is what makes the laptop demo real. At launch, Inkling 2 Mini runs on any Mac with at least 32GB of RAM through Ollama, including MacBooks and Mac Minis.

The lab’s researchers put that claim to a theatrical test. They took Inkling 2 Mini on a flight, where it ran offline on a MacBook Pro and built a playable flight simulator in a few minutes — no internet, no server, no help from the cloud.

It is a charming stunt, but it makes a serious point. A model that can run twenty persistent agents on a single laptop can also run one very capable agent for a single user — privately, offline, and without sending data to anyone.

That puts Mini in the same conversation as other local-first releases, such as Edge0-145B, a 145-billion-parameter model that runs on one consumer GPU. The difference is that Mini is smaller, faster, and built from the start for agentic work rather than raw size.

Where Inkling 2 Mini Falls Short

Thinking Machines Lab is unusually honest about the model’s limits. Inkling 2 Mini is not the strongest small model for multi-turn chat, scoring 50.0% on MultiChallenge. It also trails Qwen3.8-27B on instructions with many constraints, at 33.4% on ComplexConstraints v1.1.

It trails larger models on reasoning-heavy tests such as CritPt v2 and HLE-Diamond. Looping improves reasoning, but it cannot make the model’s weights store more facts. Knowledge, the lab notes, is mostly a function of total parameter count.

Mini compensates by using search and tools efficiently. It leads open-weight models of similar size on BrowseComp, DeepSearchQA, and MCP Atlas, which suggests it knows how to look things up rather than guessing.

Why Inkling 2 Mini Matters for Local AI

The bigger story here is not one model. It is the direction of travel. Frontier capability is quietly moving onto hardware people already own, and Inkling 2 Mini is the clearest example yet of a lab designing for that future on purpose.

A capable coding agent that runs offline changes who can use AI and how. Developers with slow internet, companies with strict data rules, and hobbyists with one good laptop all get the same tool — no subscription, no API key.

The open-weights trend is accelerating around it. Mistral Large 4 arrived as Europe’s biggest open-weight model earlier in the month. Mini shows the other end of the scale: small, fast, and good enough for real work.

There is more to come from the same lab. Inkling 2 Small will stay closed-weight briefly on Tinker, with full weights promised under Apache 2.0 next week, pending external safety testing and red-teaming. If Small follows Mini’s pattern, the open-weights frontier is about to get considerably stronger.

For now, the looped little model that built a flight simulator at 35,000 feet is free to download. The frontier, it turns out, fits in 20 gigabytes — and it runs just fine with the Wi-Fi off.

Leave a Comment

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

Scroll to Top