the biggest open AI model ever drops this week. you can download all 1.4 terabytes — you just can't run it

On July 27, Moonshot AI does something remarkable: it hands the entire world the weights to Kimi K3, a 2.8-trillion-parameter model that just took the number-one spot on the Frontend Code Arena leaderboard — ahead of Claude Fable 5, GPT-5.6 Sol and GLM-5.2. Free to download, open to run, yours to keep. There's just one catch, and it's a big one: the download is about 1.4 terabytes, and loading it needs roughly that much fast memory. 'Open weights' does not mean 'runs on your machine.' It means you can watch the file finish downloading and then discover you have nowhere to put it. Let me show you exactly why, and what you can actually do about it.

The MoE trap that fools everyone

Here's the mistake I see constantly. People hear 'Kimi K3 only activates ~50 billion parameters per token' — because it's a Mixture-of-Experts model using 16 of its 896 experts at a time — and conclude they only need enough memory for 50B parameters, which sounds almost reasonable. That's wrong, and it's the single most important thing to understand about running big MoE models. Mixture-of-Experts reduces how much compute each token needs, not how much memory the model occupies. The router can send the next token to any of the 896 experts, so all of them have to be resident in fast memory, ready. You don't get to load a tenth of the model. You load all 2.8 trillion parameters or it doesn't run.

So the memory math is unforgiving. Per the technical breakdown, Moonshot trained K3 with MXFP4 weights from the fine-tuning stage — not a crude after-the-fact compression — which gets the full weights down to about 1.4TB. At FP16 it would be 5.6TB. Even the aggressively-quantized 1.4TB figure is more memory than exists in any workstation you can buy. The good news, such as it is: the ~50B active parameters mean that if you can fit it, it runs faster than a 2.8T dense model would. The bad news is the 'if.'

So who is the open release actually for?

If almost nobody can run it locally, why open-weight a 2.8T model at all? Because 'open weights' serves people who aren't running it on a desk. Research labs studying frontier architectures like Kimi's Delta Attention and Attention Residuals. Companies that will host it on their own datacenter hardware, free of Moonshot's API terms and pricing. Cloud providers who'll offer it. And the distillation ecosystem — smaller, runnable models trained to imitate K3's behaviour, which is where most of us will actually feel its influence. The open release is a gift to the ecosystem, not to your GPU. It also, pointedly, works around US compute-export limits by letting the weights travel where the chips can't.

What you can actually run at home

Here's the payoff, because 'you can't run it' is a useless place to stop. The gap between a 2.8T frontier model and what runs on a single good GPU is smaller than the headline numbers suggest — for most real work, anyway. A Qwen3-32B-class model on a single 24GB card handles coding, analysis and chat at a level that was frontier-grade a year ago. If you want to push toward the big leagues, Llama 70B on a dual-GPU build or a big Mac is the realistic ceiling for serious home setups. And renting a datacenter GPU by the hour is how you'd touch anything K3-sized without owning a server room.

Kimi K3's open-weight drop is a genuine milestone — the largest open model ever, free to the world — and also a perfect illustration of the gap between 'open' and 'runnable.' If you want the frontier, you'll reach it through the API, a host, or the distilled models that follow. If you want something on your own hardware today, size a runnable model to your GPU in our VRAM calculator and start with what fits in 24GB. And if you're curious why K3 kept introducing itself as Claude, that's a different — and genuinely interesting — story.