expectedwrong hindsight

The Sam Altman Tweet That Wasn't About OpenAI

Kimi-K2 landed with a trillion parameters and apparently moved the most powerful man in AI to post about prices.

3 min read 543 words #open-source #kimi-k2 #moonshotai #llm #ai-pricing
hindsight — still happening

A trillion parameters, open weights, rivaling frontier models. The open-source challenge to closed labs continues escalating. The cruise ship that folds sideways is still sailing.

A trillion parameters. One trillion. Mixture-of-experts so only about 32 billion of them fire on any given forward pass, which is the polite way of saying the model is secretly enormous while pretending to be merely large — like a cruise ship that fits through the canal because it folds sideways.

Kimi-K2 dropped this week from MoonshotAI, open weights, available on OpenRouter right now, and the benchmarks say it rivals Claude Sonnet. Some say Opus. The kind of claim that a month ago would have gotten you laughed out of the room.

Nobody is laughing.

The thing about a trillion-parameter open-weights model is that it doesn't just change what you can run — it changes the negotiating position of everyone who builds on the closed ones. When a frontier-quality model is free to download and run, the conversation about API pricing becomes very different. The closed labs suddenly have to explain what they're charging for, exactly.

Sam Altman tweeted something this week. People are reading it as a response to Kimi. Maybe it was, maybe it wasn't — he doesn't typically credit the competition in his posts — but the timing is the kind of coincidence that happens when it isn't a coincidence.

That's the tell. Not the benchmarks, not the parameter count. The tell is when the CEO of OpenAI posts about prices and you don't have to squint very hard to see a Chinese lab's name in the subtext.

The pattern here is familiar if you've been watching long enough. Something releases, it's "not quite there," the serious people move on, and then six months later it's the thing everyone's using. Except Kimi-K2 is skipping the first part. It's there. It's on OpenRouter. You can hit it with an API call tonight.

Open source model releases used to feel like a charity event — here's something you can tinker with, please don't benchmark it against the real thing. That era is over and has been for a while, but Kimi-K2 is the version of "it's over" that's hard to argue with. A trillion parameters is not a hobbyist model. A trillion parameters is "we built a frontier lab and then open-sourced the output."

The agentic benchmarks are where it gets interesting. Kimi-K2 was specifically optimized for tool use and multi-step reasoning — not "chat with me about philosophy" but "go complete this SWE-bench task without a human in the loop." That's where the labs are actually competing right now. That's the real product. And Kimi lands in that specific arena and scores in the company of models that cost many multiples more per token.

The 128K context window helps. The pricing on OpenRouter is aggressive. The weights being open means someone is going to fine-tune this thing for your specific domain before the month is out, and it will be better than the general-purpose version, and it will be free.

I don't know exactly what Sam Altman said this week. But I know what Kimi-K2 said. It said: the open-source ceiling just became the closed-source floor.

Every model that wants to charge a premium now has to justify the delta between itself and something you can run yourself. That's a harder pitch than it was on Monday.