Read the Kimi K3 License Before You Read the Benchmarks

If Kimi K3 is on your model shortlist, the first file to open is not the benchmark table — it is the LICENSE. Moonshot AI made the weights of its 2.8-trillion-parameter model public on 2026-07-27, and the terms attached to them are a commercial instrument rather than a permissive open-source grant.

The weights are real, the specs are frontier-class, and the scorecard is genuinely split. But the decision this release forces is not an engineering one. We put Kimi K3 on our own shortlist on 2026-07-18 on capability and price alone; re-reading that decision against the license published on 2026-07-27 moved it out of the evaluation column and into legal review. Here is what shipped, what the license actually says, where the model wins and loses, and who should skip it.

What Shipped on 2026-07-27

The official repository at moonshotai/Kimi-K3 on Hugging Face went public with a last-modified timestamp of 2026-07-27T16:29:18Z. The repository object itself was created on 2026-06-13 and staged privately until the drop — which explains why several outlets reported the release early and why a check on 2026-07-17 correctly found no official repository on Moonshot's account.

The [model card](https://huggingface.co/moonshotai/Kimi-K3/raw/main/README.md) lists 2.8 trillion total parameters with 104 billion activated, 93 layers, 16 of 896 experts selected per token, and a 1,048,576-token context window. Attention is a hybrid: 69 Kimi Delta Attention layers plus 24 Gated MLA layers, with native multimodal input.

Moonshot's own [developer documentation](https://platform.kimi.ai/docs/guide/kimi-k3-quickstart) and its [launch post](https://www.kimi.com/blog/kimi-k3) describe it as the first open 3T-class model.

Self-hosting it is not a laptop exercise. Launch guidance points at a high-bandwidth supernode with 64 or more accelerators; quantization-aware training in MXFP4 and MXFP8 lowers the bill without changing the class of hardware you need.

The License Says "Kimi K3", Not MIT

Coverage disagreed on this point, and the disagreement matters more than the specs. One widely-read pre-release guide stated the weights would ship under a "Modified MIT license". The repository metadata says otherwise: license: other, license_name: "kimi-k3".

The [license file](https://huggingface.co/moonshotai/Kimi-K3/raw/main/LICENSE) is titled "Kimi K3 License", copyright 2026 Moonshot AI. Its opening grant is MIT-shaped — use, copy, modify, merge, publish, distribute, sublicense, sell — which is why "Modified MIT" reads plausible. Three numbered conditions after that grant are what make it something else entirely.

Open weights are not open source. As one summary of the release put it plainly, Moonshot released trained parameters, not training code or training data. The license is the second half of that distinction, and a European read of the terms lands in the same place: this text is how Moonshot intends to make money.

Three Clauses That Land on Legal, Not on Your Eval

Read verbatim from the license file, as of 2026-07-28:

Section 2 — Model as a Service. If you or your affiliates operate a business that gives third parties meaningful control over inference inputs, parameters, or training data, and aggregate revenue exceeds 20 million US dollars over any consecutive 12 months, you must enter a separate agreement with Moonshot AI before any commercial use. Embedding the model inside a specific feature, or relaying requests to models hosted by others, is explicitly excluded from that definition.
Section 3 — Attribution. Above 100 million monthly active users or 20 million US dollars in monthly revenue, "Kimi K3" must appear prominently in your product's user interface.
Section 4 — Carve-outs. Neither condition applies to purely internal use, nor to access through Moonshot's official products or certified inference partners.

The practical reading: K3 is unconditionally free for internal use and for products that embed it as a feature. It becomes a negotiation the moment you resell inference at scale. That is a procurement question, not an evaluation question, and it belongs in the same file as your other [vendor diligence](https://www.contextstudios.ai/blog/apples-openai-lawsuit-belongs-in-your-vendor-diligence) work.

Section 4 also explains why the hosted route appeared within hours. Fireworks AI and Cursor announced day-zero, US-hosted serving with Zero Data Retention. Going through a certified partner keeps you inside the carve-out and inside a US data boundary — the same framing that made hosted frontier models acceptable to enterprise buyers in the first place.

The Scorecard Is Split, and Moonshot Published Both Halves

The evaluation table on the model card is self-reported, which is the correct caveat to attach to every number below. To Moonshot's credit, it publishes losses as well as wins against Claude Fable 5 and GPT-5.6 Sol.

Benchmark	Kimi K3	Claude Fable 5	GPT-5.6 Sol
GPQA Diamond	93.5	92.6	94.1
AA-LCR (long context)	74.7	70.0	73.7
CritPt	23.4	28.6	32.3
FrontierSWE	81.2	86.6	71.3
SWE-Marathon	42.0	35.0	39.0
Terminal-Bench 2.1	88.3	88.0	88.8

The shape is consistent: knowledge, long-context retrieval and long-horizon agentic coding are strengths; hard physics-style reasoning is not. Independent indexing tells the other half of the story. Artificial Analysis ranks Kimi K3 first of 98 models on intelligence with an index score of 57, while placing it 94th of 98 on price at $3.00 per million input tokens and $15.00 per million output tokens, and 59th of 98 on speed at 32.0 output tokens per second, as of 2026-07-28.

Open weights did not make K3 cheap to consume. As of 2026-07-28 it is the most expensive way to buy the top intelligence rank, and it is slow. Self-hosting is what changes that arithmetic, and self-hosting is exactly what needs 64 accelerators.

Who Should Run It, and Who Should Skip It

Run it if you need frontier long-context work inside your own boundary, you already operate multi-accelerator inference, and your use is internal or feature-embedded. The 1M-token window plus the top long-context score is a real reason to move workloads, and it is the strongest open-weight hedge against single-vendor exposure we have written about since the US and China started gating AI differently.

Take the hosted path if you want the capability without the hardware. Certified partners keep you inside Section 4, and Zero Data Retention closes the objection most legal teams raise first.

Skip it if your product resells inference and you are anywhere near the $20M threshold, if hard reasoning benchmarks are your bottleneck, or if you were expecting an Apache-style grant you can stop thinking about. On price alone, the hosted API is not the cheap option this release was sold as.

The wider lesson is one we keep re-learning: capability is now the easy part of a model decision. Terms, data boundaries and routing are the hard part, and they are what a working agent factory has to encode. If you want help turning that into an actual routing policy across open and closed models, that is the work we do at Context Studios.

Frequently Asked Questions

Is Kimi K3 open source? No. It is open weights. Moonshot AI released the trained parameters under the Kimi K3 License, not the training code or data, and the license carries commercial conditions that no OSI-approved license contains.

Do I need a separate agreement with Moonshot AI? Only if you operate a Model-as-a-Service business — giving third parties meaningful control over inference inputs, parameters or training data — with more than $20 million in aggregate revenue over any consecutive 12 months. Internal use and feature-embedded use are exempt under Section 4.

How does Kimi K3 compare to Claude Fable 5? On Moonshot's self-reported numbers, Kimi K3 leads on GPQA Diamond (93.5 to 92.6), AA-LCR (74.7 to 70.0) and SWE-Marathon (42.0 to 35.0), and trails on CritPt (23.4 to 28.6) and FrontierSWE (81.2 to 86.6). Treat vendor-run evaluations as directional.

Can I run K3 on my own hardware? Technically yes, practically only at data-centre scale. Launch guidance recommends a high-bandwidth supernode with 64 or more accelerators; MXFP4 weights and MXFP8 activations reduce the footprint but not the class of machine.

Read the Kimi K3 License Before You Read the Benchmarks