OpenK3
Contact

Open weights guide

Kimi K3 Open Weights

Kimi K3 model weights were released on July 27, 2026 on Hugging Face. The repository contains 118 files — 96 Safetensors weight shards totalling approximately 1.56 TB, plus 4 evaluation metadata YAML files — under the custom Kimi K3 License. File size alone does not indicate whether the model can run on a given machine; official hardware requirements were not specified in the release materials. [S19, S21]

OpenK3 HomeK3 vs K2.7 CodeView on Hugging FaceTechnical Report

Release Status

Released and publicly accessible

Kimi K3 model weights were released on July 27, 2026. The repository is public — no login or access request is required. [S19, S20]

Verified against repository revision 9f62e4e at 16:45 UTC on July 27, 2026. The exact UTC release timestamp was not independently confirmed.

Official Repository

Pinned repository details

The model files are published at huggingface.co/moonshotai/Kimi-K3. The official announcement was published by the Moonshot AI WeChat account, and the technical report is available on GitHub. [S19, S20, S22]

Kimi K3 official repository status
FieldValue
Repositoryhuggingface.co/moonshotai/Kimi-K3
Revision9f62e4e9fffbd0a83ddd60e1c209d828994b3569
Last modified2026-07-27T16:29:18Z
GatedNo — public
StatusReleased

Files and Formats

Repository size and storage format

The repository contains 118 files. The 96 Safetensors weight shards are named model-00001-of-000096.safetensors through model-00096-of-000096.safetensors. Four .eval_results/*.yaml files are Hugging Face evaluation metadata; their scores are vendor-reported from the model card, not OpenK3 independent tests. [S19]

Kimi K3 repository files and formats
FieldValue
Repository total1,560,998,984,390 bytes (~1.561 TB decimal / ~1.420 TiB binary)
96 shard subtotal1,560,936,091,448 bytes (~1.561 TB decimal / ~1.420 TiB binary)
Non-weight overhead62,892,942 bytes (~62.893 MB decimal / ~59.979 MiB binary)
Weight formatSafetensors
config.json top-level dtypebfloat16
Quantizationtext_config.quantization_config: mxfp4-pack-quantized (compressed-tensors), with an ignore list excluding certain layer types
Model card quantizationMXFP4 weights / MXFP8 activations (quantization-aware training from SFT onward)

The model card also lists 2.8 trillion total parameters with 104 billion activated parameters, and a context length of 1,048,576 tokens.

OpenK3 has not downloaded this approximately 1.56 TB repository, verified shard integrity, or performed local inference.

The Model Summary table lists Text and Image. The model card's Key Features section describes video capabilities, and the hosted Kimi API documentation documents video input. A local open-weight video input workflow has not been confirmed as of this audit. [S19, S2]

License and Commercial Use

Custom Kimi K3 License

Kimi K3 uses the Kimi K3 License, a custom license with no standard SPDX identifier. OpenK3 describes this as an open-weight release because it uses the custom Kimi K3 License rather than a standard SPDX-identified license. Read the full license before use. [S21]

What the license permits

The license permits use, modification, distribution, sale, deployment, fine-tuning, and creation of derivative works, subject to copyright notice retention and compliance with applicable laws.

Key commercial conditions

  • If the Licensee or its affiliates operates a Model-as-a-Service business and aggregate revenue exceeds 20 million US dollars over any consecutive 12-month period, a separate agreement with Moonshot AI is required before any commercial use.
  • Commercial products or services with more than 100 million monthly active users or more than 20 million US dollars in monthly revenue must prominently display “Kimi K3” on the user interface.
  • The requirements in Sections 2 and 3 do not apply to internal use or to use through Moonshot AI's official products or certified inference partners.

The license is provided AS IS with no warranty.

Read the full Kimi K3 License [S21]. This summary is not legal advice.

How to Download

Use the pinned revision

The commands below point to the audited repository revision. OpenK3 has not executed these commands.

Download the pinned revision

hf download moonshotai/Kimi-K3 \
  --revision 9f62e4e9fffbd0a83ddd60e1c209d828994b3569

Preview the file list without downloading

hf download moonshotai/Kimi-K3 \
  --revision 9f62e4e9fffbd0a83ddd60e1c209d828994b3569 \
  --dry-run

Framework Support

Officially recommended inference frameworks

The official Hugging Face deployment section recommends the following frameworks. [S19]

Kimi K3 recommended inference frameworks
FrameworkOfficial example
vLLMvllm serve "moonshotai/Kimi-K3"
SGLangpython3 -m sglang.launch_server --model-path "moonshotai/Kimi-K3"
TokenSpeedListed as supported; see the official recipes page
TransformersAutoModel.from_pretrained("moonshotai/Kimi-K3", trust_remote_code=True, device_map="auto")

The official Transformers example sets trust_remote_code=True. Official minimum framework versions and patch requirements were not specified in the accessed sources. OpenK3 has not independently verified loading or inference with any framework.

Hardware Requirements

Official minimum hardware is not specified

Official GPU, VRAM, and minimum hardware requirements were not specified in the model card, the Hugging Face repository, or the release materials. [S19]

Do not estimate hardware requirements from file size, parameter counts, or quantization format. The approximately 1.56 TB repository size and mxfp4-pack-quantized storage format do not directly translate to runtime VRAM usage. Runtime memory depends on the loading library, inference engine, parallelism strategy, and runtime configuration.

Known Limitations

What remains unverified

Kimi K3 open-weight limitations and unknowns
LimitationDetail
Exact official UTC release timestampThe precise UTC timestamp of the release was not independently confirmed.
Official hardware requirementsOfficial GPU, VRAM, and minimum hardware requirements were not specified in the release materials.
Local open-weight video workflow unconfirmedThe model card mentions video; the hosted API supports video; local open-weight video support has not been verified.
Framework versions not specifiedMinimum versions and patches for vLLM, SGLang, TokenSpeed, and Transformers were not stated.
Ollama / llama.cpp status unknownSupport for these frameworks was not confirmed in the official sources reviewed.
Full repository download not executedOpenK3 has not downloaded the approximately 1.56 TB repository.
Shard integrity not independently verifiedOpenK3 has not independently verified the integrity of the 96 Safetensors weight shards.
Local inference not performedOpenK3 has not performed local inference with the open-weight repository.

FAQ

Questions about Kimi K3 open weights

Are the Kimi K3 weights released?

Yes. Kimi K3 model weights were released on July 27, 2026 on Hugging Face at huggingface.co/moonshotai/Kimi-K3. The repository is public and contains 118 files. [S19, S20]

Where can I download them?

Download them from huggingface.co/moonshotai/Kimi-K3. At the pinned revision, run: hf download moonshotai/Kimi-K3 --revision 9f62e4e9fffbd0a83ddd60e1c209d828994b3569. Add --dry-run to preview files without downloading.

How large is the repository?

The repository has 118 files, including 96 Safetensors weight shards and 4 vendor-reported evaluation metadata YAML files. Its total is approximately 1.561 TB decimal or 1.420 TiB binary. The shard subtotal is approximately 1.561 TB, with about 59.979 MiB of non-weight files. [S19]

What license covers Kimi K3?

The custom Kimi K3 License permits use, modification, distribution, and commercial deployment including derivative works, subject to conditions. OpenK3 describes this as an open-weight release because the license has no standard SPDX identifier. Read the full license before use. [S21]

Can I use Kimi K3 commercially?

Commercial use is permitted subject to the license conditions. If the Licensee or its affiliates operates a Model-as-a-Service business with aggregate revenue exceeding 20 million US dollars over any consecutive 12 months, a separate agreement with Moonshot AI is required. Products with more than 100 million monthly active users or 20 million US dollars in monthly revenue must prominently display “Kimi K3.” The requirements in Sections 2 and 3 do not apply to internal use or to use through Moonshot AI's official products or certified inference partners. Read the full license for complete terms. [S21]

What hardware is required?

Official GPU, VRAM, and minimum hardware requirements were not specified in the model card or release materials. The approximately 1.56 TB repository size and mxfp4-pack-quantized storage format do not directly indicate runtime VRAM usage. [S19]

Official Sources

Documents used for this guide

Each source below includes its own access date.

  1. [S1]
    Kimi K3: Open Frontier Intelligence — Moonshot AI official blogAccessed July 24, 2026https://www.kimi.com/blog/kimi-k3
  2. [S2]
    Kimi K3 Quickstart Guide — Kimi API documentationAccessed July 24, 2026https://platform.kimi.com/docs/guide/kimi-k3-quickstart
  3. [S19]
    Moonshot AI — Kimi K3 Hugging Face Model RepositoryAccessed July 27, 2026 at 16:45 UTChttps://huggingface.co/moonshotai/Kimi-K3
  4. [S20]
    月之暗面 Kimi official WeChat release announcementAccessed July 27, 2026https://mp.weixin.qq.com/s/tryHe81IyM6nr0fBPDz72g
  5. [S21]
    Kimi K3 LicenseAccessed July 27, 2026https://huggingface.co/moonshotai/Kimi-K3/blob/9f62e4e9fffbd0a83ddd60e1c209d828994b3569/LICENSE
  6. [S22]
    MoonshotAI/Kimi-K3 Technical Report RepositoryAccessed July 27, 2026https://github.com/MoonshotAI/Kimi-K3