What the open-weight release is
Moonshot AI released the Kimi K3 model weights on July 27, 2026 through the public moonshotai/Kimi-K3 repository on Hugging Face. The repository is not gated, so no login or access request is required to view or download it.
OpenK3 audited the repository at pinned revision 9f62e4e9fffbd0a83ddd60e1c209d828994b3569, last modified 2026-07-27T16:29:18Z. The weights are distributed under the custom Kimi K3 License, which has no standard SPDX identifier; OpenK3 therefore describes this as an open-weight release rather than assigning it a standard open-source license label.
Repository snapshot at the pinned revision
The audited revision contains 118 files totalling 1,560,998,984,390 bytes — approximately 1.561 TB decimal or 1.420 TiB binary. The core of the repository is 96 Safetensors weight shards, with the config.json top-level dtype listed as bfloat16.
The model card lists 2.8 trillion total parameters with 104 billion activated parameters and a context length of 1,048,576 tokens. Four .eval_results YAML files are Hugging Face evaluation metadata; their scores are vendor-reported from the model card, not OpenK3 independent tests.
Download and license in brief
The official hf download command can target the pinned revision directly, and adding --dry-run previews the file list without transferring data. Kimi K3 download file structure and pinned checks walks through the inventory and the verification boundaries in detail.
The Kimi K3 License permits use, modification, distribution, sale, deployment, fine-tuning, and derivative works, subject to conditions that include revenue and user thresholds for some commercial deployments. Kimi K3 License summary explains those conditions and what remains unknown. Read the full license before use.
Framework support and hardware unknowns
Official materials reference vLLM, SGLang, TokenSpeed, and Transformers examples. The official Transformers example sets trust_remote_code=True. Minimum framework versions and patch requirements were not specified in the accessed sources, and Ollama and llama.cpp support was not confirmed.
Official GPU, VRAM, and minimum hardware requirements were not specified in the model card, the repository, or the release materials. The approximately 1.56 TB repository size and the mxfp4-pack-quantized storage format do not directly translate to runtime VRAM usage, which depends on the loading library, inference engine, parallelism strategy, and runtime configuration.
What this audit does not cover
OpenK3 has not downloaded the approximately 1.56 TB repository, verified the integrity of the 96 shards, or performed local inference. A local open-weight video input workflow has not been confirmed, even though the model card and the hosted API describe video capabilities.
This overview is descriptive only. It makes no hardware, compatibility, or benchmark conclusions, and the vendor-reported evaluation metadata should be read as Moonshot AI's own reported results rather than independent measurements.