New Free AI Thinks 58% Less and Runs Twice as Fast

An AI research team called UkisAI has released a free, tweaked version of an open-source AI model that does something unusual: it thinks less. When modern AI answers a question, it first writes a long stretch of internal reasoning — roughly like showing its work. Sometimes it gets stuck going in circles. UkisAI trained its version to avoid those repetitive loops, cutting the thinking by 58% and making it run about twice as fast, with less than 1% loss in answer accuracy.

Why should you care? Thinking costs money and time. Every word an AI writes internally burns computing power, which is why AI services charge per use and often feel slow. Nearly halving that work means the same answers for roughly half the cost and double the speed. The model is small enough — 27 billion parameters — that hobbyists and small companies can run it on a decent laptop or a modest rented server instead of a giant data center.

The catch: this is a mid-sized, open-source model, not a frontier system like the ones behind ChatGPT or Gemini. It won't match them on hard problems. The free trial API is limited to five requests per minute and is for research use only. And the team is upfront that this doesn't replace simply asking an AI to think less — it fixes a separate problem, the random loops models fall into.

UkisAI also published compressed versions so the model fits on smaller machines, and the community has already made uncensored and low-precision variants. The team credits Nvidia for donating the GPUs used to train and host it. In short: a modest but genuinely useful efficiency win, handed out for free — the kind of thing that quietly makes everyday AI cheaper.