What it takes to run the biggest open models

The weights are free to download. The hardware is not. Kimi-K3 has 2780B parameters and needs about 1575 GB of memory at 4-bit precision — roughly 22 H100 cards, or $594,000 of silicon before you have powered any of it on.

Nobody is running these at a desk — that isn't the point. It's worth seeing the scale of what “open” actually means at the top end, and how far it sits from the models on the rest of this section. What your own computer can run →

The models, by size

Two prices, because there are two ways to do this. 13 of these 22 fit a single 512GB Mac Studio at about $9,500 — one machine, no server rack. The H100 column is what a datacentre build costs, and you pay that for speed rather than capacity: 3,350 GB/s per card against the Mac's 819, multiplied across every card.

Memory is at 4-bit with 4K context — the smallest anyone sensibly runs these at, and at full precision they are roughly four times larger again. Models that ship natively quantised are measured from their published weights rather than a conversion. Card counts divide memory by the H100's 80GB with a 10% allowance, so they are a floor: real deployments lose more to activations, long-context caches and sharding. Both prices are indicative street prices for the hardware alone, excluding the server, networking, power and cooling an H100 build needs and a Mac does not.

Kimi-K3 on different hardware

The same model, and what it would take on each generation of accelerator.

Speeds assume a single card's bandwidth and are optimistic: splitting a model across several cards adds communication cost that a single-card figure ignores.

Can't you just quantize it down until it fits?

Not far enough. Quantization has a floor, and these models start so far above it that even the most extreme settings land well beyond any consumer machine. DeepSeek is the example people ask about most, so here are the real published file sizes:

Measured from the published GGUF conversions at unsloth/DeepSeek-V3.1-GGUF. DeepSeek R1 goes slightly lower — its smallest is 140 GB — but that still needs a 192GB machine once the context cache and runtime are counted.

Squeezing a model to one bit per weight is not free either. By that point it has lost enough precision that it makes mistakes the full version wouldn't, so you are running a noticeably worse model on hardware that still cost more than a car.

So how is everyone running DeepSeek?

Three ways, and only one of them is a laptop:

• Through an API. Overwhelmingly the most common. The weights being public doesn't mean people host them themselves — it means anyone may, and a handful of providers do.
• On a distilled version. This is the one that causes the confusion. Repos like DeepSeek-R1-Distill-Qwen-32B and DeepSeek-R1-0528-Qwen3-8B run happily on a laptop and have millions of downloads — but they are Qwen models trained on DeepSeek's output, not DeepSeek's own architecture. When somebody says they're running DeepSeek locally, it is usually one of these.
• On serious hardware. A 256GB or 512GB Mac Studio, or a multi-GPU rig, running a heavily quantized build slowly. Real, but rare, and expensive.

The same is true of every model on this page. “Open weights” is a licensing fact, not a hardware one.

What about GPT, Claude and Gemini?

We can't tell you. OpenAI, Anthropic and Google don't publish parameter counts or release weights for their frontier models, so any figure for them would be a guess dressed up as a fact. Every number on this page comes from a published file size or an official specification, and we'd rather leave a gap than fill it with invention.

That difference is itself the point of the page. With an open-weight model you can measure exactly what it costs to run, because the weights are right there. With a closed one you get a price per token and no way to check what sits behind it.