Thinking Machines Lab Launches Inkling 2 Mini — Open-Weights AI That Runs on Your Laptop

Thinking Machines Lab has released Inkling 2 Mini, a fast, powerful model that is small enough to run on your laptop. Announced on November 20, it is a multimodal open-weights model built for local hardware — and its headline trick is looping: the middle layers of the model run three times for every token, so a 26-billion-parameter model behaves like a much bigger one.
The release matters because it answers a question that keeps coming up in local AI: can a model that fits on one machine actually do serious agentic work? Thinking Machines Lab’s answer is a demo with a flight simulator built offline on a MacBook Pro, and benchmark numbers that beat the current local favourite.
The looping trick, in plain terms
Inkling 2 Mini is a mixture-of-experts transformer with 26.4 billion total parameters and 3.4 billion active per token. It has 24 decoder layers arranged in four blocks of six. The first block runs once, the two middle blocks run three times each, and the final block runs once — so each token passes through 48 layer applications even though only 24 layers are stored. Each token also selects 8 of 128 experts in a layer, and the model carries 5.4 billion weight activations per token once looping is counted.
That looping is what lets a small checkpoint punch above its size. The model was trained natively in 4-bit precision — 20 layers in NVFP4, the last four in MXFP8 — and the released checkpoint keeps that exact layout, so there is no conversion loss from training precision to deployment precision. The total memory footprint is about 20GB, which is why it fits on consumer hardware at all.
What it can do, and where it wins
Inkling 2 Mini accepts text and image inputs and generates text, with a context window of up to 1 million tokens. Its strengths mirror those of its bigger sibling, Inkling 2 Small: long-horizon agentic work, coding assistants, chatbots, tool use and retrieval-augmented generation.
On benchmarks, it leads every open-weights model with less than 100 billion total parameters on DeepSWE v1.1, scoring 48.2%, and on Vals Index v2.2 at 46.5%. In Thinking Machines Lab’s own comparisons it beats Qwen3.8-27B — the model most local coding setups default to today — on DeepSWE, SWE-bench Pro v2 Hard, SWE-bench Multilingual, SWE-bench Verified and Vibe Code Bench. Those are vendor-reported numbers, but they are direct head-to-head comparisons against the same model, which makes them far more useful than launch-post cherry-picks.
Stanford researchers measured it another way: intelligence per joule, or how much energy a model needs per query to reach a given level of performance. Inkling 2 Mini scored 0.703 running on an NVIDIA DGX Spark, against 0.430 for Qwen3.6-35B-A3B, 0.159 for Gemma-4-26B-A4B and 0.036 for Qwen3.5-27B. For anyone running agents all day on local hardware, that efficiency gap is the number that actually shows up in the electricity bill.
Running it on your own machine
The model is released with open weights under the Apache 2.0 licence, on Hugging Face, with a companion research model (TML-d24-varloop-research) alongside it. Inference and fine-tuning are also available through Thinking Machines Lab’s Tinker platform, with a playground for trying it without downloading anything.
For local inference, the official hardware targets are a single NVIDIA DGX Spark, or Apple silicon devices with 32GB or more of memory — MacBooks, Mac minis and Studios. Other GPUs and platforms are possible with 32GB or more of VRAM recommended. Supported frameworks include Hugging Face transformers, SGLang, vLLM, TokenSpeed, NVIDIA Dynamo, Unsloth and llama.cpp for quantized GGUF versions, plus Ollama and Exo Labs on Apple silicon and the Llama App.
The demos show why that matters. Researchers took Inkling 2 Mini on a flight, where it ran offline on a MacBook Pro in airplane mode and built a playable flight simulator in a few minutes. Another demo, “Thinkyville”, is a tiny pixel town where the same model powers all 20 villagers on one offline MacBook — a stand-in for running multiple persistent agents on local hardware, with no cloud round-trip at all.
Why local models keep getting better
Inkling 2 Mini lands in a month when the big labs have been pushing the other direction — Gemini 4 Neon as a fast mid-tier cloud model, and ever-larger data-centre hardware like NVIDIA’s Rubin HPC GPU. The local track is the counterweight: your data stays on your machine, your agents keep working on a plane, and your costs stop scaling with every token.
That is also why the open-weights licence matters. Developers can fine-tune it, build it into products and run it inside their own tools without asking permission or paying per token. If you have a 32GB Mac or a DGX Spark-class machine, Inkling 2 Mini is now the local model to beat — and the download is free on Hugging Face today.
