Kimi K3 Open Weights: What Shipped and How the Repository Is Actually Organized

Moonshot AI released the Kimi K3 weights on Hugging Face last week, and the repo immediately became one of the largest open-weight checkpoints most of us have ever seen. If you're trying to figure out what exactly shipped and how the files fit together, the official repo is the source of truth — but it's also ~1.5 TB and 100+ files, so orienting yourself takes time.

I put together an independent reference called OpenK3 that maps out the current Hugging Face revision, the file layout, and the license summary. This post walks through what I learned while building it.

What actually shipped

The moonshotai/Kimi-K3 repo contains the full weights for a 2.8T-parameter sparse MoE model. At the time of writing, the checkpoint is organized around 96 safetensors shards plus the usual supporting files: config.json, tokenizer files, model index mappings, and evaluation metadata.

Key points if you're planning to download:

• The full checkpoint is roughly 1.56 TB of safetensors.
• It uses MXFP4 compression, so you'll need serving software that can handle it (vLLM and SGLang have day-0 or near-day-0 support, with caveats depending on your GPU generation).
• The repo also includes .eval_results/*.yaml files that describe the reported benchmark metadata. These are not independent test results — they're the evaluation configuration/outputs Moonshot published alongside the model.

Repository layout at a glance

If you've only worked with smaller models before, the file tree can be surprising. Besides the shard files, you'll see:

• config.json — architecture and quantization settings
• tokenizer.json / tokenizer_config.json — the tokenizer and special tokens
• model.safetensors.index.json — maps logical layer names to shard files
• generation_config.json — default generation parameters
• .eval_results/*.yaml — reported eval metadata
• License files

I wrote up the layout in more detail here: Kimi K3 open weights guide.

Download with revision pinning

Because the repo is large, you don't want to discover halfway through that the upstream revision changed. Pin the revision when you clone:

huggingface-cli download moonshotai/Kimi-K3 \
  --revision 9f62e4e9fffbd0a83ddd60e1c209d828994b3569 \
  --local-dir ./Kimi-K3

That revision matches the layout I documented. If you see a newer revision, double-check the file count and structure before treating any guide as current.

Before you self-host, read the license

Kimi K3 is released under a custom license, not a standard OSI-approved license. Commercial use is allowed, but there are revenue-tier conditions and prohibited use cases. I summarized the commercial-use terms in plain language here, but it's an independent reference — not legal advice, and not a substitute for reading the license yourself: Kimi K3 license summary.

Bottom line

The K3 release is a big deal for open-weight frontier models, but the practical barrier isn't just hardware — it's also knowing what you're downloading and what you're allowed to do with it. Use the official Moonshot AI docs and repo as the primary source, and use OpenK3 as a secondary map if you want a quicker orientation.

If you've already downloaded the checkpoint, I'd love to hear what serving stack worked for you in the comments.