Kimi K3 is a 2.8-trillion-parameter AI model. You still can't run it — even with 8 GPUs

Moonshot AI just shipped the largest open-weight AI model ever released — a 2.8 trillion-parameter system called Kimi K3 — and it beat Anthropic's Claude Fable 5 on a head-to-head coding benchmark. That's the headline. The number underneath it is the one that actually matters if you've ever priced out a GPU: at 2.8 trillion parameters, this model does not fit in anything a serious hobbyist owns, and the math on why is worth doing in full.

One caveat up front: Moonshot hasn't shipped the full weights yet — those are due July 27. But the architecture (total parameters, expert count, activation pattern) is already public, and that's all the hardware math below actually needs.

The mixture-of-experts trick — and why it doesn't save you

K3 is a mixture-of-experts (MoE) model: instead of running all 2.8 trillion parameters on every token, it routes each token through just 16 of its 896 experts — about 1.8% of the total. That's why Moonshot can serve it cheaply through an API: the compute cost per token is closer to a much smaller dense model. But compute and memory are different budgets. The router doesn't know which 16 experts a token needs until it sees the token, so every one of those 896 experts has to be sitting in fast memory, ready, on every single request. MoE saves you FLOPs. It does not save you a single byte of VRAM.

The actual math

1.4 TB is the optimistic number. An Nvidia H100 80GB card — the workhorse of actual AI data centers, not a consumer product — gets you there in roughly 18 cards, wired together with the kind of high-bandwidth interconnect that doesn't exist on a desktop motherboard. Eight top-end RTX 5090s, the most VRAM-per-dollar a consumer can buy right now, add up to 256 GB combined: not even a fifth of what K3 needs at its most compressed. This isn't a model you under-provision for. It's a model that was never a local-hardware story to begin with.

"Open-weight" tells you what you're legally allowed to do with the file. It doesn't tell you whether you can afford to load it.

What actually changed for people running models at home

Nothing, directly — and that's the real story. K3 is part of a pattern: the frontier keeps scaling past what any single enthusiast rig can hold, while the models people actually run at home — dense 7B–70B checkpoints, or smaller sparse models sized for one or two GPUs — keep improving on a completely separate track. Moonshot's own numbers show K3 still trailing Claude Fable 5 and GPT-5.6 Sol on general capability; it wins one specific, narrow coding benchmark. For a home rig, the more useful question was never "is there a bigger model," it's "what's the best model that fits the card I can actually buy" — and that answer moves on its own schedule, mostly unrelated to headlines like this one.

We track that second question here: which frontier techniques trickle down into models sized for one to four consumer GPUs, and which stay API-only by design. If you're speccing a local-AI build, the VRAM math above is the whole game — it just isn't K3's game.