Image: Thinking Machines Lab
SAN FRANCISCO, November 21, 2026 — Thinking Machines Lab, the AI start-up founded by former OpenAI chief technology officer Mira Murati, has released Inkling 2 Mini, an open-weights model it says is fast and capable enough for serious coding work yet small enough to run on a laptop.
In an announcement dated November 20, the company said the full model has a memory footprint of about 20GB, allowing it to run on consumer laptops and desktop AI computers. Its researchers said they ran it offline on a MacBook Pro during a flight, where it built a playable flight simulator in a few minutes.
What is a “looped” model?
Most language models pass each word, or token, through their layers once. Inkling 2 Mini is a looped transformer: it runs its middle layers three times for every token. Think of it as re-reading a hard sentence before answering. Thinking Machines says this lets the model reason more and reach the right answer with fewer tokens, especially on agentic coding tasks, where an AI works through multi-step jobs on its own.
It is also a Mixture-of-Experts (MoE) model, meaning only part of the network switches on for each token. It has 26.4 billion parameters in total, with 3.4 billion active, or 5.4 billion weight activations per token once looping is counted. Most of its weights were trained and are served in NVFP4, a compact 4-bit number format that keeps memory needs low.
How it performs
The company says Inkling 2 Mini leads every open-weights model under 100 billion total parameters on benchmarks such as DeepSWE v1.1 (48.2 per cent) and Vals Index v2.2 (46.5 per cent). It was candid about the limits: looping improves reasoning but does not let the model store more facts, so its factual-knowledge score is in line with models of similar size. It compensates by using search and tools.
Thinking Machines also released a research variant with a variable number of loops and said the model’s chain of thought grew shorter with more loops but stayed readable, which matters for anyone trying to monitor what an AI is doing.
The model is available with open weights on Hugging Face and on the company’s Tinker platform.
What to watch
For Malaysian developers and businesses, capable models that run locally could cut cloud costs and keep sensitive data on their own machines. Watch for independent benchmark results, and whether looped designs catch on among other open-model makers.
Source: Thinking Machines Lab. Featured image: Thinking Machines Lab.
