I love when one model launch flips the whole conversation on its head. Last week Kimi K3 dropped from a Chinese lab called Moonshot, and suddenly people are arguing about the future of the entire AI industry. This breakdown comes from an AI creator who walked through exactly why this release matters, and I couldn’t stop taking notes.
Here’s the short version he lays out. Kimi K3 is a 2.8 trillion parameter open-source model with a 1 million token context window, and it’s genuinely comparable to the best closed models like GPT 5.6 and Fable. The wild part? A Chinese company built a frontier-level model and is giving the recipe away for free.
Wait, what does open source even mean here?
The author explains it cleanly. Closed models like ChatGPT and Claude are fully controlled by their makers, from pricing to who gets access. Open-source models get released to the world, including the data, the algorithms, and the training techniques. The US leans closed. China leans open, with Deepseek, Moonshot’s Kimi, and Alibaba’s Quen all handing their work out for free.
So why give it away?
This was my favorite part of his explanation. Companies open source the layer they want to commoditize so they can win somewhere else. He breaks down the reasons China does it:
- Ecosystem control: become the default standard everyone builds on
- Scorched earth: crush competitors’ margins by making AI nearly free
- State policy: China subsidizes these labs as a geopolitical play
He points out this playbook is old and proven. Linux runs the world’s servers. Android runs most phones. Chromium shaped web browsing. React powers frontends. Open source wins more than people think.
Who actually wins?
The creator pulls in a sharp argument from investor Gavin Baker. A world with only two or three dominant labs holding 90% inference margins is bad for every other layer, chips, energy, data centers, and especially startups. Cheaper tokens mean more usage, which means more demand across the whole stack. That’s Jevons paradox at work. The only real loser is the closed frontier lab.
The cost twist most people miss
The author references a Ben Thompson piece and adds nuance I appreciated. Kimi looks half the price of GPT 5.6, roughly “$3 per million input tokens and $15 per million output tokens.” But Kimi uses about double the tokens to reach the same answer. So the metric that matters is cost per task, not sticker price. Not every token is equally smart, so tokens aren’t a true commodity.
Why is the US talking about a ban?
He walks through an Axios report suggesting the Trump administration might restrict cutting-edge Chinese models. His take is that a ban is neither realistic nor helpful. He even cites David Sacks noting Kimi K3 fixed 15 critical security bugs that guardrailed US models refused, and Hugging Face’s CEO describing the same defensive problem. The fear is that squeezing open source just hands the market to OpenAI and Anthropic.
My takeaway
The person who put this together lands on a clear position: allow open source, stay cautious with Chinese models, and compete at every layer instead of hiding behind bans. He’d rather see closed labs ship faster and own responsibility through know-your-customer checks. Honestly, I found that framing hard to argue with.
There’s way more in the full video, including the distillation attack drama between Anthropic and these Chinese labs. Go watch the whole breakdown, it’s worth your time. 🚀