Kimi K3: A Free-to-Download Model That Almost Keeps Up With the Big Names
China's Moonshot released Kimi K3, an open-weight model that lands just behind Fable 5 and GPT-5.6 Sol in independent tests. It's also a lot pricier than the last one, a sign the era of dirt-cheap Chinese AI is ending.
Chinese lab Moonshot has released K3, its new flagship AI model, and this one is a bit of a milestone. It’s an open-weight model (meaning the finished model can be downloaded and run by anyone, not just used through the company’s website) in the roughly 3-trillion-parameter range. In plain terms: it’s one of the largest freely available models anyone has shipped, and it performs close to the best paid models on the market.
Here are the numbers, minus the launch-day glitter. K3 has 2.8 trillion total parameters and a context window of one million tokens (roughly, how much text it can keep in mind at once, here about a very long book’s worth). The independent testing lab Artificial Analysis scored it 57 on its Intelligence Index, putting it fourth overall behind Claude Fable 5 (60), GPT-5.6 Sol (59), and Claude Opus 4.8 (56). On agent-style tasks, the kind where a model works through many steps on its own, it beat GPT-5.5 and Opus 4.8 while still trailing Fable 5. The full weights are due by July 27.
There’s a catch worth knowing before you get excited: K3 makes things up more often than its predecessor. Its hallucination rate (how often it states something false with confidence) climbed to 51 percent in one test, even as it got more answers right overall. So it’s smarter and more confident, which is a tricky combination if you don’t check its work.
The part the benchmarks do not shout about is the price. K3 costs about $3 per million input tokens and $15 per million output, roughly the same as Anthropic’s Sonnet 5 and far above the near-free rates Chinese labs offered a year ago. The old Kimi K2.6 was many times cheaper. The takeaway: Chinese providers are no longer racing to the bottom on price. They now charge like serious Western mid-range products, because they think their models finally justify it.
What this means for you: If you don’t build software, nothing changes today, but the direction matters. Strong open models push the paid ones to stay cheap and useful. If you do tinker with AI tools, K3 is worth a look once the weights land, especially for long coding jobs, just don’t trust it on facts without double-checking. And the bigger story for everyone: “cheap Chinese AI” is quietly becoming “competitive Chinese AI at a real price.”
Sources
OpenAI Made a $230 Keypad Just for Bossing Around AI Coding Agents
OpenAI and keyboard maker Work Louder released the Codex Micro, a small desk controller with a joystick, dial, and light-up keys for managing AI coding agents. It's niche, but it hints at where AI work is heading: its own dedicated hardware.
