New Free AI Models Cut the Overthinking, Answer Twice as Fast

An independent developer called UkisAI just released a family of free AI models it calls Swift. These aren't products you sign up for — they're files anyone can download and run on their own computer. They're built on top of Qwen, an open AI model from Alibaba that anyone can modify. The previous Swift version was downloaded over 350,000 times in under two weeks, which is a lot for a no-name project. This release includes three updated models: Swift1.5 27B, Flash Next, and Bonsai 2.

Here's the idea. Many modern AI models "reason" out loud before answering — writing pages of internal notes, second-guessing themselves, sometimes going in circles. That's genuinely useful for hard math problems, but it's slow and it costs money, because most commercial AI is billed by how much text it processes. Think of a colleague who narrates every thought before giving you an answer versus one who just answers. UkisAI trained its models to skip the pointless spiraling. Flash Next, for example, used 63% fewer reasoning steps and ran about 1.8 times faster, while scoring within a rounding error of the original.

The catch: all these numbers come from UkisAI's own testing, and the differences are tiny — fractions of a percent, which is close to statistical noise. One headline test score actually looks worse at first glance, which the team explains as the model stubbornly finishing tasks rather than giving up, and burning more words doing so. These are also experimental tools aimed at hobbyists and tinkerers, not plug-and-play apps. If you just use ChatGPT or Claude, nothing changes for you today.

So why care? Because this is the direction everything is heading. Big AI companies are under enormous pressure to make reasoning cheaper, and open-source experiments like these are often where the tricks first appear. Cheaper reasoning eventually means cheaper AI subscriptions, faster assistants, and the ability to run a capable AI on a laptop without an internet connection.