You can download it. You can't run it

Yesterday I covered Qwen3.8-Max and said that even when the weights drop, almost nobody will run a 2.4 trillion parameter model themselves. Today I wanted to prove it with real numbers, using a model you CAN already download: Kimi K3, Moonshot AI's 2.8 trillion parameter open-weight flagship with a 1 million token context. I priced out what self-hosting it actually takes, just for the memes and the giggles, and the answer explains a lot about what "open weights" really means.

Why self-host at all

Kimi K3 is genuinely good, I've been testing it on my own agent workflows for the complicated tasks my local models can't handle, and it's cheap compared to the frontier closed models. But Moonshot AI is a Chinese company, and China's National Intelligence Law requires Chinese companies to cooperate with government intelligence requests. For regulated industries, finance, healthcare, legal, calling a mainland API is an immediate no-go. If you want this class of model AND full control of your data, self-hosting is the clean answer. So what does that take?

The ladder, tier by tier

The full-precision model is 5.6 terabytes. That's the download. Running it is a VRAM problem, and the ladder looks like this:

Reference quality (full precision): 5,600+ GB of VRAM, roughly 100 H100 80GB GPUs. This is "same quality as Moonshot's own hosting."
Near-lossless (Q8): about 2,800 GB, roughly 35 H100s.
The sweet spot (Q4): roughly 20 GPUs. K3 was trained quantization-aware (MXFP4), so the Q4 quality drop is smaller than usual, this is what most providers will actually serve.
The big squeeze (Q2): around 900 GB of VRAM, with meaningful quality loss. Even the most compressed version of this model needs a rack, not a PC.

H100s run $27,000 to $40,000 each depending on new versus used. Do the ballpark: 120 cards at an average $25k is $3 million in GPUs alone, before power, cooling, networking, storage, or the building they live in. The Q4 tier still lands around $500K. This is datacenter money, not homelab money.

The middle paths

If you don't want to own hardware, AWS will rent you the equivalent: three p5.48xlarge instances (8x H100 each) at roughly $100 an hour on demand per instance, less reserved. That only pencils out if your API bill is already enormous.

The realistic answer for almost everyone: managed hosting on Western infrastructure. Providers like Together AI and Fireworks already serve Kimi K3 on US infrastructure at the same price as Moonshot's own API, $3 per million input tokens, $15 out, with zero-retention data policies. I checked live on OpenRouter during the video: US region, same price. You get the model, US data residency, and no cluster to manage. The pricing is identical whether you go to mainland China or a US host, which honestly surprised me.

One open question: the K3 license still hasn't been confirmed (DeepSeek, for comparison, uses MIT weights). Until it lands, "open" has an asterisk here too, same as Qwen3.8.

What this means

The point isn't that open weights are fake. It's that for frontier-sized models, "open" mostly benefits companies with infrastructure, and shows up for the rest of us as price pressure and provider choice, both real wins. Me personally? My money's on the smaller siblings. My local Qwen models handle my day-to-day private workflows on hardware I own, and that's the version of "open" that actually reaches regular people. The 2.8 trillion parameter one? I priced it. That was the fun part.

Sources

All prices checked August 4, 2026.