Kimi K3 hardware requirements: the open weights are 1.56 TB, and here's what that means for your GPU

Kimi K3 is a real open-weight release. Moonshot put the full model on Hugging Face, under its own Kimi K3 License, and anyone can download it. Running it is another matter.

I read the model card, the config file, the license and the file list on 4 October 2026, plus vLLM's serving recipe and Moonshot's launch blog. I have not run Kimi K3 on any hardware, and nobody at home has the hardware to. Everything below is arithmetic on published files.

How much memory does Kimi K3 need?

About 1.56 TB for the official weights, before any context cache or runtime overhead. Moonshot ships K3 in MXFP4, a 4-bit format it trained with from the fine-tuning stage on, so there is no bigger "real" version you are missing. Community quants go lower, but the smallest is still 509 GB.

The model card lists 2.8T total parameters, 104B activated, and "MXFP4 weights / MXFP8 activations (quantization-aware training)." Here is what each download weighs.

The bits-per-weight column is derived: file size times 8, divided by the 2,779,931,837,184 parameters Hugging Face counts in the official repo. It runs above 4.0 for the MXFP4 files because some layers, such as attention and the shared experts, are kept at higher precision. The config lists them in its "ignore" list.

What hardware does Kimi K3 actually run on?

Datacenter nodes. vLLM's Kimi K3 recipe lists "At least 8x GB300," with "Multi-node for real production traffic," and at least 8 MI355X or MI350X GPUs on AMD. Moonshot's own blog goes further and recommends "supernode configurations with 64 or more accelerators." That is rack-scale hardware.

The vLLM recipe's launch command runs the model with tensor parallel size 8, an fp8 context cache and the full 1,048,576-token window. If you want to touch K3 without owning that, the honest options are Moonshot's API or a cloud provider that hosts it. Our rent vs buy guide covers when renting makes sense at all.

Why doesn't "104B active" make it fit?

Because active parameters set compute, not memory. K3 picks 16 of its 896 experts for every token, plus 2 shared experts. The next token can go to any expert, so all 2.8 trillion parameters must sit in fast memory. You pay memory for the whole model and compute for about 104 billion.

Older write-ups, including an earlier version of this page, put the active count near 50B. The model card says 104B. That higher figure makes no difference to whether it fits. It does mean each token costs about twice the compute those estimates assumed.

This is the trap with every giant mixture-of-experts model. Offloading experts to system RAM, the --n-cpu-moe trick, works when a model is a bit too big for your GPU. It cannot bridge 509 GB on a desktop with 64 or 128 GB of RAM.

The cache side is lighter than you might fear. Only the 24 Gated MLA layers keep a classic context cache, and MLA stores a compressed version of it. The other 69 layers use linear attention with a fixed state. So context adds memory, but the weights are the wall.

Is Kimi K3's license really open?

Mostly. The Kimi K3 License grants the usual rights to use, modify, host and fine-tune. Two conditions apply only at scale: a model-as-a-service business with over $20 million revenue in 12 months needs a separate agreement, and products past 100 million monthly users or $20 million monthly revenue must display "Kimi K3."

Internal use is exempt from both conditions, per section 4 of the license. For research labs and most companies, that is effectively open. As always, read the LICENSE file before you build on it. This is a summary, not legal advice.

What can I run at home instead of Kimi K3?

A lot, for most real work. A 24 GB card runs a 27B to 32B model at Q4 with long context. A 64 GB or larger unified-memory machine, or one GPU plus plenty of RAM, runs gpt-oss-120b. And 12 to 16 GB cards have their own MoE options. None match K3, but all fit.

The two home-scale rows come from our VRAM engine and the official file sizes, read 4 October 2026. See the gpt-oss-120b hardware requirements for that model, the cost to run Qwen3.8 27B, and our list of the best MoE models for 12 to 16 GB GPUs. For a single 24 GB card, start with what fits in 24 GB.

Kimi K3 is a milestone for open weights and a lesson in the gap between "open" and "runnable." Reach it through an API or a host. Run something sized to your card at home, and check the fit first in our VRAM calculator. If you are curious why K3 kept introducing itself as Claude, that is a different story.