Kimi K3 Open Weights: Download & Self-Hosting Guide
On July 27, 2026, Moonshot AI released the full Kimi K3 weights on Hugging Face — making K3 the largest open-weight model ever published at 2.8 trillion parameters. This guide covers how to download it, what you can realistically run it on, what the license allows, and the honest question most people should ask first: should you self-host at all?
Where to download
The official repository is moonshotai/Kimi-K3 on Hugging Face. It crossed 99,000 downloads on release day.
The fastest way to download is the Hugging Face CLI:
pip install huggingface_hub
huggingface-cli download moonshotai/Kimi-K3Make sure you have enough free disk space before starting — this is a multi-hundred-GB model, and you'll want download resume support. Check the model card for the exact file list and sizes.
What you're actually getting
- Architecture: Mixture-of-Experts, 2.8T total parameters, 104B active per token (16 of 896 experts)
- Context: 1,048,576 tokens (1M)
- Multimodal: native text + image + video understanding (MoonViT-V2 encoder)
- Quantization: MXFP4 weights / MXFP8 activations, trained in from the start — this is what makes a 2.8T model feasible to serve at all
- License: the Kimi K3 License (a modified license, not plain MIT) — read it on the model card before commercial use
The hardware reality check
Let's be direct: you are not running this on your laptop, and not on a single consumer GPU either. Even quantized, a 2.8T-parameter model needs a serious multi-GPU server setup. Self-hosting K3 makes sense for:
- Companies with data-residency or compliance requirements
- Teams that want to fine-tune the model on their own data
- Inference providers and researchers with existing GPU clusters
If you're an individual developer who just wants to use K3, self-hosting is the wrong path. Use one of these instead:
- Free chat: kimi.com — no setup, no cost
- Official API: platform.kimi.ai — model name
kimi-k3, OpenAI-compatible, $3/$15 per 1M tokens (input/output) - OpenRouter: K3 on OpenRouter — if you want one unified API key across providers
If you do self-host: supported inference engines
Moonshot recommends these engines for K3 (recipes linked from the model card):
- vLLM — the most common choice for serving open models
- SGLang
- TokenSpeed
Community runtimes like Ollama may follow, but at launch the officially supported path is the three engines above. Watch the model card and Moonshot's GitHub for updates — deployment recipes for new releases tend to improve rapidly in the first weeks.
Bottom line
K3's open weights are a big deal for the ecosystem: frontier-tier intelligence you can actually download. But "open-weight" doesn't mean "runs anywhere" — for 95% of people, the right way to use K3 today is the free chat or the API. Self-host only if you have the hardware and a concrete reason.
Sources
moonshotai/Kimi-K3 model card · Kimi API documentation · Links checked Jul 27, 2026. K3 Box is an independent directory, not affiliated with Moonshot AI.