Kimi K3 Went Open at 1.56 Terabytes

What actually landed on Hugging Face

Moonshot AI published the weights for Kimi K3 on Hugging Face, and this time the files are genuinely there. The repository carries 96 safetensors shards, a config.json and a LICENSE file, and the listing puts the total at 1.56 terabytes. That last number matters more than it looks, because the weights are stored at MXFP4 precision with MXFP8 activations. There is no smaller version waiting to be quantised: this is the compressed distribution.

The architecture is a mixture of experts at 2.8 trillion total parameters with 104 billion activated per token, drawn from 896 experts of which 16 are selected each time. The stack runs 93 layers, 69 of them KDA and 24 gated MLA, on a 160,000 token vocabulary with a default context window of 1,048,576 tokens. It is a vision-capable model, listed as image-text-to-text rather than text only.

The published scores are strong across the categories that matter to working teams: 93.5 on GPQA Diamond, 88.3 on Terminal-Bench 2.1, 91.2 on BrowseComp, 67.5 on DeepSWE and 81.6 to 83.4 on MMMU-Pro for vision. Moonshot states that both the code repository and the model weights are released under the Kimi K3 License, which is where the interesting part begins.

Two thresholds, and the smaller-looking one is the binding one

The Kimi K3 License is not Apache and it is not MIT. It carries two separate commercial triggers, and coverage will almost certainly conflate them. The first sits in section three: if the software or a derivative is used in a commercial product or service with more than 100 million monthly active users, or more than 20 million dollars in monthly revenue, then the words Kimi K3 must be prominently displayed in that product's interface. That is an attribution duty, it applies to a very short list of companies, and it does not restrict what you may build.

The second sits in section two and works differently in every respect. If you or any of your affiliates operate a Model as a Service business, and the aggregate revenue of you and your affiliates exceeds 20 million dollars in total over any consecutive twelve months, you must enter into a separate agreement with Moonshot AI before using the software commercially at all. Not a logo. A signed contract, and a condition precedent to use.

Do the arithmetic on the two numbers, because they are not the same size. Twenty million dollars per month is 240 million a year. Twenty million dollars across twelve months is one twelfth of that trigger, roughly 18 million euros of group revenue at current rates. A profitable European software company with fifty staff can sit above the clause that gates use while sitting far below the clause that gates branding. The threshold that reads as the lenient one is the one that actually reaches into the mid-market.

Open weights are not a runnable model at 1.56 terabytes

Set the licence aside for a moment and look at the logistics. Pulling 1.56 terabytes is not a developer afternoon; it is a scheduled transfer with egress implications, and it lands you a model that activates 104 billion parameters per token and expects serious accelerator memory to serve at a sensible latency. The open publication is real and it matters for auditability and for research, but it does not put this model inside most companies.

What that means in practice is that the majority of European firms who end up using Kimi K3 will reach it through somebody else's infrastructure: a neocloud, a hyperscaler endpoint, or an inference provider. The word open will be doing very little work in that arrangement. You will be buying tokens from a vendor, exactly as you do today with closed models, and the openness will show up as a price argument rather than as control.

That is not a criticism of Moonshot. It is the standard consequence of frontier scale, and it applies to every large open-weight release. The mistake to avoid is treating published weights as evidence that you have an exit. An exit requires somewhere to go, and 1.56 terabytes of MXFP4 shards is a destination very few balance sheets can host.

What to settle before anyone points an application at it

Section four of the licence contains the clause most readers will actually rely on. The commercial requirements do not apply to internal use, nor to access through Moonshot's own official products or its certified inference partners. That exemption is genuinely useful, and it is also a relocation of the problem: if your route to the model is a host, then whether you are compliant is determined by your host's status and by what your contract with that host says. The model card cannot answer it for you.

Three things to do before deployment. Establish your own group revenue figure against the twelve-month test and write down the answer, because the clause counts affiliates and most groups have not computed that number for this purpose. Ask counsel whether anything you sell constitutes a Model as a Service business, since reselling inference in any wrapper is the exposure and product teams rarely think of themselves that way. Then, if you are going through a provider, require written confirmation that it is a certified inference partner and that the exemption flows to you.

Do all three before code ships, because section two is a condition on using the software commercially rather than a penalty applied afterwards. Discovering it during a diligence review is considerably more expensive than discovering it now.

Read next: A Frontier Model Just Stopped Taking Customers  |  Independent Tests Rank Kimi K3 Above Fable 5