ToolNavs Find Useful AI Tools
Submit Sign in
Back to AI Q&A
How to Run Inkling 2 Mini on a Laptop: For a 26B Model, Work Out Memory and Licensing First

How to Run Inkling 2 Mini on a Laptop: For a 26B Model, Work Out Memory and Licensing First

AI Q&A • Admin • • 3 views

How to run Inkling 2 Mini on a laptop comes down to working out memory and licensing first: Thinking Machines Lab released this 26B-parameter looped-architecture model on November 20, 2026. It has 26.4B parameters in total and activates only 3.4B at a time, its weights take up about 20GB, and the company says it can run offline on a laptop. The weights are open under the Apache 2.0 license — download and go — but 26B is not the same as lightweight. Without enough memory it still won't run, so do the sums before you start.

Do the memory sums first

The weights alone are about 20GB, and then you need room for runtime caches and framework overhead — the official demo ran on a 48GB MacBook Pro, the one with the M5 Pro chip. By that measure, 32GB of memory is a reasonably safe starting point; a 16GB machine will most likely need more aggressive quantization or put up with a crawl. Don't look at the 3.4B active parameters and assume it's small — that's a measure of compute, not a memory budget.

Two ways to get it running on a laptop

If you want the easy route, use a one-click tool like Ollama: install it, pull the model, and start chatting — good for trying it out, for chat, and for simple agent tasks. If you want performance for agent work and long context, deploy the full weights with vLLM or SGLang — vLLM already supported it on launch day, and Ollama is on the supported list too. The weights are available from Thinking Machines Lab's official model page on Hugging Face, and there's also a GGUF quantized build for llama.cpp.

Licensing and where it fits

The Apache 2.0 license allows commercial use, modification, and redistribution, as long as you keep the copyright notice and license text — and there are no subscription fees. The model takes text and image input with a 1M-token context window, and the company says it stands out on long-horizon agent work. But it has weak spots too: on multi-turn chat and instructions packed with constraints, it trails same-size models such as Qwen3.8-27B. Before making it your main model, run it on your own tasks first.

Q: Where's the safest place to download the weights?

A: Thinking Machines Lab's official model page on Hugging Face — look for the name thinkingmachines/Inkling-2-Mini, and don't download weight files of unclear origin from unknown sources.

Recommended Tools

More