aliteq.

Type to search every guide, tool and GPU page.

Kimi K3's open weights are 1.56 TB. Can your GPU actually run it?

Moonshot's Kimi K3 is open to download. It is also 2.8 trillion parameters in 1.56 TB of MXFP4 files, and the smallest community quant is 509 GB. The real memory math, what vLLM says you need, and what to run at home instead.

Ari ValeUpdated Published 8 min read
Kimi K3 'Open Frontier Intelligence' announcement page from Moonshot AI

Some links on this page are referral links: we may earn a commission, at no cost to you.

Share

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: the official weights are 1.56 TB and the smallest community quant is 509 GB, so it needs a rack of datacenter GPUs, not a home PC or a gaming rig.

I read the model card, the config file, the license and the file list on October 4, 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.

Kimi K3: size of each download (read Oct 4, 2026)

MXFP4 (official)

Total size
1,561 GB (1,454 GiB)
Bits per weight
about 4.5
Source
Moonshot AI

NVFP4

Total size
1,610 GB (1,499 GiB)
Bits per weight
about 4.6
Source
NVIDIA

GGUF UD-Q4_K_XL

Total size
1,509 GB (1,405 GiB)
Bits per weight
about 4.3
Source
unsloth, community

GGUF UD-Q2_K_XL

Total size
861 GB (802 GiB)
Bits per weight
about 2.5
Source
unsloth, community

GGUF UD-IQ1_S

Total size
594 GB (553 GiB)
Bits per weight
about 1.7
Source
unsloth, community

GGUF UD-TQ1_0 (smallest)

Total size
509 GB (474 GiB)
Bits per weight
about 1.5
Source
unsloth, community

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.

Does it fit? Kimi K3 official MXFP4 needs about 1,561 GB
Kimi K3 (MXFP4) needs ≈1561 GB
RTX 5090 (32 GB)32 GBover 1529 GB
4x RTX PRO 6000 (384 GB)384 GBover 1177 GB
8x H200 (1,128 GB)1128 GBover 433 GB
Weights only. Even eight H200s (141 GB each) fall short of the official files before any context is added.
The official Moonshot AI organisation page on Hugging Face, host of the Kimi model weights
The weights sit on Moonshot AI's official Hugging Face org. Free to download, far beyond any home machine to run. · Hugging Face (screenshot)

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.

Could you rent enough GPUs yourself? The official files are out of reach even for eight H200s, which hold 1,128 GB between them. The 509 GB UD-TQ1_0 quant is a different story: four H200s hold 564 GB, enough for the weights plus a modest context. On Vast.ai an H200 started at $3.75 an hour on Nov 17, 2026, so a four-card machine runs roughly $15 an hour. At about 1.5 bits per weight that is a test drive of K3, not the full model. Our H200 rental cost guide compares hosts, and the live cloud GPU price table shows today's rates.

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.

Kimi K3 architecture (model card and config.json, read Oct 4, 2026)

Total parameters

Value
2.8T

Active per token

Value
104B

Layers

Value
93: 69 Kimi Delta Attention + 24 Gated MLA

Experts

Value
896, with 16 selected + 2 shared per token

Context

Value
1,048,576 tokens

Vision encoder

Value
MoonViT-V2, 401M parameters

Weight format

Value
MXFP4 weights, MXFP8 activations (quantization-aware training)

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 such as a used RTX 3090 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. Even a base Mac mini can run the smaller gpt-oss-20b, with the limits laid out in whether gpt-oss-20b runs on a 16GB Mac mini.

Kimi K3 vs models that fit at home

Kimi K3 (MXFP4)

Memory needed
about 1,561 GB
Runs on
8+ datacenter GPUs, multi-node for production

gpt-oss-120b (MXFP4)

Memory needed
about 61 GiB weights
Runs on
64 GB+ unified memory, or a GPU plus RAM offload

Qwen3.8 27B (Q4_K_M, 32K context)

Memory needed
about 18.5 GiB
Runs on
one 24 GB card

The two home-scale rows come from our VRAM engine and the official file sizes, read October 4, 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.

No 64 GB machine either? gpt-oss-120b's 61 GiB of weights fit on one 80 GB H100 with no offload, which is the closest thing to a big open model you can rent by the hour without a cluster. Our cheapest H100 rental guide compares where that card costs least.

If the hardware bill is out of reach, the Kimi K3 API from Moonshot is pay as you go: $3.00 per million input tokens and $15.00 per million output (read Oct 29 and re-checked Nov 12, 2026).

This guide is part of our Which local AI model fits my GPU's VRAM? The 8GB to 32GB chart, model by model guide and our What hardware do DeepSeek guide. Related guides: Meta promised to open-source Muse Spark 1.2, a model that rivals GPT and Claude, What's the best local LLM for a 32GB RTX 5090, Best local LLM for 8GB VRAM: what fits (moved to our fuller guide).

Kimi K3 hardware: quick answers

Can I run Kimi K3 on an RTX 5090 or two RTX 3090s?
No. The smallest community quant is 509 GB and the official weights are 1,561 GB. A 32 GB RTX 5090 or 48 GB across two 3090s holds well under a tenth of the smallest file.
How many GPUs does Kimi K3 need?
vLLM's recipe asks for at least 8 GB300 GPUs, or at least 8 MI355X or MI350X on AMD, and multi-node setups for production traffic. Moonshot recommends supernodes with 64 or more accelerators.
Is there a Kimi K3 GGUF?
Yes, community GGUFs exist, for example unsloth's set from 509 GB (UD-TQ1_0) to 1,509 GB (UD-Q4_K_XL). Even the smallest needs about 474 GiB of fast memory before any context.
How many parameters are active in Kimi K3?
The model card lists 104B activated out of 2.8T, from 16 of 896 experts plus 2 shared experts per token. That cuts compute per token, not memory.
Can I use Kimi K3 commercially?
Generally yes under the Kimi K3 License. A model-as-a-service business over $20 million in yearly revenue needs a separate agreement, and very large products must display "Kimi K3." Internal use is exempt from both.

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.

Found this useful? Share it

Share

Founder & chief editor

Ari Vale

Founder and chief editor of Aliteq (a PC hardware shop from 2019, now a US-based publication). Ari Vale is my pen name: I stay anonymous, the work doesn't. I research the questions people actually ask, check every price and limit against the vendor's own pages, and write down what I'd choose and why. I use the cloud-GPU and AI-video services I write about.

Work out the hardware

GPUs in this article

Specs, the AI models each one runs, and where to buy or rent it.

Email list

Local AI, only the changes that matter

A new model your GPU can actually run, a GPU price move worth acting on, or a fix that saves you an evening. Nothing in between.

Written and sent by hand. One-click unsubscribe in every email. Emails may contain referral links.

Keep reading