File inventory at the pinned revision
At revision 9f62e4e9fffbd0a83ddd60e1c209d828994b3569, the moonshotai/Kimi-K3 repository contains 118 files totalling 1,560,998,984,390 bytes — approximately 1.561 TB decimal or 1.420 TiB binary. The repository is public and not gated.
The main weight set consists of 96 Safetensors shards named model-00001-of-000096.safetensors through model-00096-of-000096.safetensors, with a subtotal of 1,560,936,091,448 bytes. The remaining non-weight overhead is 62,892,942 bytes, approximately 62.893 MB decimal or 59.979 MiB binary.
Formats, dtype, and quantization metadata
The config.json top-level dtype is bfloat16. The configuration lists text_config.quantization_config as mxfp4-pack-quantized using compressed-tensors, with an ignore list excluding certain layer types. The model card describes MXFP4 weights and MXFP8 activations from quantization-aware training starting at the SFT stage.
Four .eval_results YAML files are Hugging Face evaluation metadata whose values originate from the model card. They are vendor-reported scores, not independent OpenK3 tests, and they should be labelled that way whenever they are cited.
Pinned download and dry-run preview
Download the pinned revision with: hf download moonshotai/Kimi-K3 --revision 9f62e4e9fffbd0a83ddd60e1c209d828994b3569. Pinning the revision ensures the command requests exactly the snapshot OpenK3 audited rather than whatever the default branch points to later.
Add --dry-run to the same command to preview the file list without transferring anything. A pinned command identifies the requested files; it does not establish sufficient storage, memory, network capacity, or a working inference environment.
Checks to keep separate from file size
Official GPU, VRAM, and minimum hardware requirements were not specified in the reviewed release materials. Repository size and quantization metadata do not directly determine runtime VRAM, which depends on the loading library, inference engine, parallelism strategy, and runtime configuration.
Official materials reference vLLM, SGLang, TokenSpeed, and Transformers examples, but minimum framework versions were not specified, and Ollama and llama.cpp support was not confirmed in the official sources reviewed. The Kimi K3 License conditions apply to distribution and deployment, so read the license before sharing or serving the weights.
Open verification items
OpenK3 has not downloaded the approximately 1.56 TB repository, so transfer duration, resumed-download behavior, and post-download integrity of all 96 shards remain unverified by OpenK3. Local inference has not been performed with any framework.
A local open-weight video input workflow has also not been confirmed. Treat these as open items to verify against your own environment rather than as facts this article can settle.