Forget bigger models. The hottest trend in AI right now is making them think less. A small outfit called UkisAI just dropped a family of free, downloadable models that slash the internal "thinking tokens" that normally make reasoning AIs slow and expensive. We're talking cuts of up to 63.4% — like a marathon runner shaving two hours off their time without losing speed. The best part? The benchmarks barely budge.
Take Swift-Qwen3.8-27B. It reduces thinking tokens by 58.3% and runs nearly twice as fast (x1.95 speedup) with less than 1% accuracy loss. On GPQA-Diamond, it scores 88.4% vs. 88.3% for the base model — a rounding error. On LiveCodeBench v6, it actually jumps from 76.8% to 81.6%, a +4.8pp gain, while cutting median thinking tokens by 46%. That's like a student who writes half the essay and gets a better grade.
But here's the twist: the most aggressive model, Swift 1.5 Qwen3.8-Flash-Next, trims 63.4% of thinking tokens and speeds up 1.8x with only a -0.2% accuracy dip. And then there's Swift Bonsai 2 — the weird one. It cuts thinking by 39.8% and scores 0.19% *higher*. UkisAI itself calls it experimental. That's either a breakthrough or a glitch. Either way, it's fascinating.
Why should you care? Because every time you chat with an AI, it's doodling in the margins — writing out long internal monologues before answering. That's slow and expensive. If you're paying by the word, this is your tax cut. For enterprises building voice agents (like OpenAI's new Presence tool) or robotics startups (like those in Google DeepMind's accelerator), faster, cheaper reasoning could be the difference between a demo and a product.
Of course, these numbers come from secondary sources — aggregators, a YouTube demo, an HN thread. No official blog post or model card. So take the headline stats with a grain of salt. But the direction is clear: the AI industry's obsession with brute-force thinking is being challenged by a scrappy team that says, "Maybe you don't need to think so hard."
If this trend holds, the giants — OpenAI, Google, Anthropic — might have to answer a simple question: why are your models still overthinking? The next six months could see a race to the bottom on reasoning tokens, not just benchmark scores. And that could be the best thing to happen to your AI bill in years.
So watch this space. Because if UkisAI can make a 27B model almost twice as fast with no accuracy loss, the big labs might suddenly find themselves overthinking their own strategies.