Kimi K3's weights are genuinely open — published on 27 July 2026, eleven days after the API launch. But they are not MIT, and the difference matters if you are building a business on top of them. The Kimi K3 License carries specific revenue and user thresholds that most coverage skips.
What you get
| Released | 27 July 2026 |
| Format | MXFP4 weights, MXFP8 activations |
| Size | ~1.4 TB |
| Licence | Kimi K3 License |
| Serving engines | vLLM, SGLang, and TokenSpeed |
This is a complete frontier model, not a distilled or reduced variant, and the quantised form is the trained artefact rather than a post-hoc compression.
The two thresholds you need to know
Model-as-a-Service revenue. If you operate a MaaS business on these weights and your revenue exceeds $20M over any consecutive twelve months, you must negotiate a separate commercial agreement with Moonshot AI. Reselling inference at scale is not covered by the open licence.
Attribution at scale. If your product exceeds 100 million monthly active users or $20M in monthly revenue, "Kimi K3" must be displayed prominently in your user interface.
Internal use, and access through Moonshot's own products or certified partners, are exempt from these restrictions.
How this compares to MIT
MIT imposes no thresholds, no attribution beyond the copyright notice, and no obligation to negotiate anything with anyone. The Kimi K3 License is more restrictive on all three counts.
For the overwhelming majority of users — internal tools, products below the thresholds, research, evaluation — the practical difference is nil. For anyone planning to resell inference, or building something that might plausibly reach a hundred million users, it is a conversation to have with counsel before you architect around it. The software is also provided as-is, with no warranty.
Why open weights matter even if you never self-host
Most people reading this will use the API. The weights still change your position in three concrete ways:
- Pricing leverage. A model you could run yourself constrains how far its price can move.
- Continuity. If the API is deprecated or the terms change, the weights you downloaded still work.
- Auditability. Regulated environments can inspect and evaluate the artefact directly rather than trusting a description of it.
That is the real argument for caring about this release even at $3.00 per million tokens on someone else's hardware.
More on Kimi K3
Start with the complete Kimi K3 guide for the overview, or go deeper:
Ready to go deeper?
Read the full Kimi K3 guideFrequently Asked Questions
Is Kimi K3 open source?
It has open weights, which is not the same as open source. They are published under a custom Kimi K3 License rather than MIT or Apache. Commercial use is broadly permitted, but there are revenue and user thresholds that trigger additional obligations.
Can I use Kimi K3 commercially?
Yes, with two conditions. Model-as-a-Service operators whose revenue exceeds $20M over any consecutive 12 months must negotiate a separate agreement with Moonshot AI. Products above 100 million monthly active users or $20M monthly revenue must display "Kimi K3" prominently in the interface. Internal use is exempt.
Where can I download the Kimi K3 weights?
They were published on Hugging Face on 27 July 2026 under the moonshotai organisation, at roughly 1.4 TB in MXFP4 format. Serving is supported by vLLM, SGLang, and TokenSpeed. Check the hardware requirements before downloading — this is a multi-node deployment.
