Where do I download the Kimi K3 open weights?

Kimi K3's full open weights live at huggingface.co/moonshotai/Kimi-K3 as MXFP4 safetensors shards, alongside a matching MoonshotAI/Kimi-K3 GitHub repo that ships the tokenizer, tech report, and reference inference code. Community BF16 and GGUF re-quants are expected on Hugging Face within days of the July 27 drop.

What license does Kimi K3 use?

Kimi K3 ships under a Modified MIT license that Moonshot brands as the 'Kimi K3 License,' allowing research, commercial deployment, and further fine-tuning of the weights. This is the same permissive family Moonshot used for Kimi K2.6, and it makes Kimi K3 the largest permissively licensed open-weight AI model available anywhere.

What hardware do I need to self-host Kimi K3?

Third-party self-hosting guides put the minimum for Kimi K3 at 8× H100 80 GB or equivalent, using vLLM, SGLang, or TokenSpeed as the inference engine. A single consumer GPU cannot load Kimi K3 even in the native MXFP4 format — the ~594 GB weight file only fits on multi-GPU servers or with disk-backed offload.

How does open-weight Kimi K3 compare to Claude Fable 5 and GPT-5.6 Sol?

Kimi K3's official model card reports GPQA Diamond 93.5 and DeepSWE 67.5, and arena front-end coding tests show developers picking Kimi K3 ahead of prior US frontier models. It still trails Claude Fable 5 and GPT-5.6 Sol on overall rankings, but Nathan Lambert's Interconnects analysis argues the open-to-closed gap has shrunk from 6–9 months to 3–5.

Why is Moonshot releasing Kimi K3 open-weight now?

Moonshot's Kimi K3 open-weights drop is timed to land just ahead of the 2026 World Artificial Intelligence Conference in Shanghai and follows a run where Moonshot's revenue grew at least 6× and ARR climbed from $200M in April to $300M in June. Moonshot frames the release as making 'frontier intelligence openly available for research, deployment, and further innovation.'

Moonshot AI · 2026-07-27 · major

Kimi K3 open weights — Moonshot drops the 2.8T MoE on Hugging Face

Moonshot AI published the full open weights for Kimi K3 on Hugging Face at 00:00 UTC on July 27, making the 2.8-trillion-parameter mixture-of-experts model the largest open-weight AI release in history and free to self-host.

Hugging Face social banner for the moonshotai/Kimi-K3 model repository

The largest open-weight AI model in history — 2.8T Kimi K3 is now free to self-host.

Key specs

Active params	104B
Context window	1M tokens
Total parameters	2.8T
Download size	~594 GB (MXFP4)

Quick facts

Maker	Moonshot AI (China)
Architecture	Mixture-of-Experts — 896 experts, 16 active per token
Vision encoder	MoonViT-V2 (401M params)
Context window	1,048,576 tokens (1M)
Quantization	MXFP4 weights, MXFP8 activations
License	Modified MIT (Kimi K3 License)
Download size	~594 GB across safetensors shards
Minimum hardware	8× H100 80 GB (per third-party guides)

What is it?

Kimi K3 is Moonshot AI's flagship mixture-of-experts model with 2.8 trillion total parameters and 104 billion active per token, and on July 27 its full weights went live on Hugging Face for anyone to download and run. That makes Kimi K3 the largest permissively licensed open-weight AI model ever released, and the first 3T-class model that any lab or research team can self-host.

How does it work?

Under the hood, Kimi K3 uses 896 experts with 16 activated per token and a 1-million-token context window. MoonViT-V2, a 401M-parameter vision encoder, ships in the same weights so the model handles text, images, and video natively. Moonshot quantizes the release to MXFP4 weights and MXFP8 activations, bringing the download to about 594 GB and letting vLLM, SGLang, and TokenSpeed serve it on multi-H100 nodes.

Why does it matter?

Self-hosting Kimi K3 gives teams that could never send data to a Chinese API the same 2.8T-parameter reasoning and 1M-token context inside their own network. It also resets the open-versus-closed race: Nathan Lambert argues in Interconnects that the frontier gap has narrowed from 6–9 months down to 3–5, and the Kimi K3 release lands the same day Anthropic publishes its formal position on open-weight risk.

Who is it for?

AI labs, sovereign-compute buyers, and self-hosters with 8+ H100s

Frequently asked questions

Where do I download the Kimi K3 open weights?: Kimi K3's full open weights live at huggingface.co/moonshotai/Kimi-K3 as MXFP4 safetensors shards, alongside a matching MoonshotAI/Kimi-K3 GitHub repo that ships the tokenizer, tech report, and reference inference code. Community BF16 and GGUF re-quants are expected on Hugging Face within days of the July 27 drop.
What license does Kimi K3 use?: Kimi K3 ships under a Modified MIT license that Moonshot brands as the 'Kimi K3 License,' allowing research, commercial deployment, and further fine-tuning of the weights. This is the same permissive family Moonshot used for Kimi K2.6, and it makes Kimi K3 the largest permissively licensed open-weight AI model available anywhere.
What hardware do I need to self-host Kimi K3?: Third-party self-hosting guides put the minimum for Kimi K3 at 8× H100 80 GB or equivalent, using vLLM, SGLang, or TokenSpeed as the inference engine. A single consumer GPU cannot load Kimi K3 even in the native MXFP4 format — the ~594 GB weight file only fits on multi-GPU servers or with disk-backed offload.
How does open-weight Kimi K3 compare to Claude Fable 5 and GPT-5.6 Sol?: Kimi K3's official model card reports GPQA Diamond 93.5 and DeepSWE 67.5, and arena front-end coding tests show developers picking Kimi K3 ahead of prior US frontier models. It still trails Claude Fable 5 and GPT-5.6 Sol on overall rankings, but Nathan Lambert's Interconnects analysis argues the open-to-closed gap has shrunk from 6–9 months to 3–5.
Why is Moonshot releasing Kimi K3 open-weight now?: Moonshot's Kimi K3 open-weights drop is timed to land just ahead of the 2026 World Artificial Intelligence Conference in Shanghai and follows a run where Moonshot's revenue grew at least 6× and ARR climbed from $200M in April to $300M in June. Moonshot frames the release as making 'frontier intelligence openly available for research, deployment, and further innovation.'

Try it

huggingface-cli download moonshotai/Kimi-K3