Kimi K3 torrent: download the 2.8T open weights

Hugging Face was crawling the morning Kimi K3 dropped. That’s what happens when a 1.5 TB repo goes up and everyone tries to clone it at once. Someone in a Discord I lurk in posted a .torrent file instead, which downloaded a lot faster.

I checked it against the official repo before trusting it. It’s the real thing, so I’m mirroring it here.

The Kimi K3 torrent links

Torrent file: Kimi-K3.torrent (1.8 MB)

Infohash (BTv1 / SHA-1): 1e63a865fbf9b58decc8b71091db54d673c5da6f

Magnet link, if you’d rather skip the file:

magnet:?xt=urn:btih:1e63a865fbf9b58decc8b71091db54d673c5da6f&dn=Kimi-K3&tr=udp%3A%2F%2Ftracker.opentrackr.org%3A1337%2Fannounce&tr=udp%3A%2F%2Ftracker.openbittorrent.com%3A6969%2Fannounce&tr=udp%3A%2F%2Fopen.stealth.si%3A80%2Fannounce&tr=udp%3A%2F%2Fexplodie.org%3A6969%2Fannounce&tr=udp%3A%2F%2Ftracker.torrent.eu.org%3A451%2Fannounce&tr=udp%3A%2F%2Fexodus.desync.com%3A6969%2Fannounce

You get 114 files, 1,560,998,983,621 bytes in total. That’s 1.56 TB, or 1.42 TiB if your client counts that way. Ninety-six of those files are safetensors shards. The other 18 are the shard index, the config, the modeling code, the tokenizer, and the model card. It’s commit c5d1dd4 of moonshotai/Kimi-K3, copied exactly.

Hugging Face is listed as a web seed, which matters more than it sounds. If nobody’s seeding, your client pulls the missing pieces over HTTP instead of sitting at 3% forever. You don’t have to switch anything on for this, qBittorrent and Transmission both use web seeds automatically when a torrent has them.

But grab the .torrent file rather than the magnet if you want that. Web seeds live outside the part of a torrent that magnet links can carry, so a magnet gets you the files and the trackers but not the Hugging Face fallback.

You’ll also see about 826 MB of padding files. Every real file is aligned to a 16 MiB piece boundary, so no piece ever spans two files. It’s the considerate way to build a torrent this big.

What’s actually in the 1.5 TB

Kimi K3 is Moonshot’s 2.8T-parameter mixture-of-experts model with 104B active parameters.

Two numbers in config.json surprised me. First, it fires 16 experts out of 896. Kimi K2 fired 8 out of 384, so K3 has more than twice as many experts and each one is smaller. Second, most of the attention isn’t full attention at all. Only 24 of the 93 layers run gated MLA. The other 69 are Kimi Delta Attention, which is how they got to a million tokens of context without the KV cache blowing up.

The weights ship in MXFP4 straight from training, not quantized afterward. Moonshot says quantization-aware training started at the SFT stage, with MXFP8 activations.

Not everything is 4-bit, though. The config lists what stays at higher precision: self-attention, the shared experts, the dense MLP projections, lm_head, and the whole vision tower. Those are the parts where losing precision hurts most, so it’s a sensible split.

No, you cannot run this on a mini PC

Let’s do the math, because someone always asks.

You need 1.5 TB of fast memory just to hold the weights, before the KV cache. That means a rack of H200s or MI355Xs. Even the Strix Halo boxes I keep recommending for local LLMs max out at 128 GB, so you’d need twelve of them wired together perfectly. The interconnect alone would turn it into a very expensive space heater.

If you want something you can actually run, Gemma 4 on a CPU-only mini PC is where that ends. Moonshot recommends vLLM, SGLang, and TokenSpeed for K3, and there’s a hosted API if you just want to use it.

So why download it at all? Same reason I keep my own services on hardware I control. Right now most people’s copy of a frontier model lives on one company’s CDN, and that’s fine until it isn’t. A torrent is a cheap way to have a second copy.

If you’re going to download it anyway

A few things worth knowing before you start:

• Check your disk first: 1.42 TiB free, plus room for partial files. Running out of space at 90% on a transfer this size hurts.
• Use the .torrent rather than the magnet, so you get the Hugging Face web seed.
• You probably don’t want all 96 shards. If you only came to read the modeling code, deselect the safetensors in your client. That’s a 63 MB download instead of 1.5 TB.
• Seed it afterward if you have the bandwidth. Mirrors only help while somebody keeps them up.

Read the license before you build on it

The Kimi K3 License is Moonshot’s own, not Apache or MIT, and it has two conditions that are easy to miss.

If you run a Model as a Service business and you and your affiliates pull more than $20 million in total revenue over any 12 consecutive months, you need a separate agreement with Moonshot before using K3 commercially. Their definition of Model as a Service is giving third parties access to inference or fine-tuning in a way that lets them control the inputs, parameters, or training data. Embedding the model inside a feature of your own product doesn’t count, and neither does relaying requests to someone else’s hosted copy.

The second one is attribution. Ship a product using K3 with more than 100 million monthly active users, or more than $20 million in monthly revenue, and you have to display “Kimi K3” prominently in the interface.

Neither applies to internal use, which the license defines as anything that doesn’t expose the model, its outputs, or its capabilities to third parties. So tinkering at home is fine. If you’re anywhere near those numbers, read the actual text rather than my summary of it.

Stuff people keep asking

Is this the official torrent? No. Moonshot publishes the weights on Hugging Face and ModelScope. This is a community mirror of the exact same files.

Is downloading it legal? Moonshot published these weights themselves. Sharing them over BitTorrent is passing around files that are already free to download. What you’re allowed to do with them afterward is the license’s business, and there’s a section above on that.

How long does 1.5 TB take? About three and a half hours on a gigabit line if nothing gets in the way. Realistically, start it before bed.

Can I check the download afterward? Your client already did. That’s what the piece hashes are for. If you want to double-check, compare each shard’s SHA-256 against the lfs.oid values in the Hugging Face tree API.

Why is it only 1.5 TB if it’s 2.8T parameters? Four-bit weights. MXFP4 plus its group scales works out to 4.25 bits per weight, and the tensors held at higher precision push the real average to about 4.5. That lands right around 1.5 TB. The same model at bf16 would be 5.6 TB.

Resources

• Kimi K3 on Hugging Face, the official weights and model card
• Moonshot’s K3 tech blog for the architecture writeup
• vLLM, one of the three inference engines Moonshot recommends
• Best mini PC for local LLMs in 2026 for models you can actually run
• Run Gemma 4 on a mini PC, no GPU required, CPU-only inference that fits in a drawer

Twelve Strix Halo boxes. I did price it out. My wife has seen the spreadsheet and the answer is no.

So it’s sitting on a spare drive in the basement instead, seeding. See you in the swarm. 🌱

Last updated: July 2026.