Kimi K3 Open Weights Are Live: The License Catch

Kimi K3's Open Weights Are Out. Read the License First.

Moonshot AI published the full open weights for Kimi K3 on Hugging Face on July 27, 2026. That makes it the first open-weight model to cross the 3-trillion-parameter class, and it changes the self-hosting math for any team weighing a frontier model against Claude. It also comes with a license that Moonshot itself refuses to call "open source," and the distinction is not just semantics.

This post covers what actually shipped, what the license lets you do and where it stops you, and how K3 stacks up against Claude Opus 5 and Claude Fable 5 for teams deciding whether to self-host or stay on an API.

What Moonshot Actually Released

Kimi K3 is a 2.8 trillion parameter mixture-of-experts model with 104 billion parameters activated per token, a 1-million-token context window, and native support for text, image, and video input in the same model, according to Moonshot's own model card on Hugging Face. It runs on a new attention architecture Moonshot calls Kimi Delta Attention (KDA) paired with Attention Residuals, and selects 16 of 896 experts per token.

The weights ship in native MXFP4 quantization, a 4-bit format Moonshot trained the model in from the fine-tuning stage onward rather than compressing after the fact. The Hugging Face repository holding the full weights runs to roughly 1.56 terabytes across 96 shards, per independent coverage of the release, and running it at production speed means real infrastructure once you account for activation overhead on top of the raw weights. Either way, this is not a model you run on a single workstation GPU.

Kimi K3 at a glance:

Open Weight, Not Open Source. Here's the Difference That Matters.

Moonshot calls K3 "open-weight" throughout its own documentation and never uses the term "open source." That word choice is deliberate, and independent commentators covering the release, including developer Simon Willison, noted Moonshot is being consistent about it rather than blurring the line the way some vendors do.

The distinction shows up directly in the license text. Here is what the Kimi K3 License, published alongside the weights on Hugging Face, actually says:

• You can use, copy, modify, fine-tune, distribute, and deploy the model commercially at no license fee, for the vast majority of use cases. The core grant is broad and MIT-derived.
• If you run a "Model as a Service" business, meaning you give a third party API-style access to K3's inference or fine-tuning with meaningful control over inputs or parameters, and your aggregate revenue with affiliates exceeds $20 million over any rolling 12-month period, you must sign a separate commercial agreement with Moonshot AI before continuing. Reselling raw model access at scale is the one path this license does not leave open by default.
• If your product built on K3 crosses 100 million monthly active users, or $20 million in monthly revenue, you have to display "Kimi K3" prominently in that product's user interface. That's an attribution requirement most teams will never hit, but if you're building toward that scale, plan for it now rather than discovering it later.
• Internal use is fully exempt from both clauses, as is any use routed through Moonshot's own official products or certified inference partners.

None of this stops a typical engineering team from fine-tuning K3 for an internal tool, building a product around it, or shipping it commercially. What it stops is quietly reselling raw K3 access as a hosted API business once you clear real revenue, without talking to Moonshot first. Read the full license text before you commit engineering time to a K3-based product, not after.

What This Means for Self-Hosting a Frontier Model

Open weights change the conversation from "which API do we call" to "do we want to run this ourselves." That's a real option now for teams with the infrastructure to back it, but it's not a free option.

A ~1.56TB weight repository does not fit on a single GPU, or a single server in most cases. Running K3 at anything close to its intended speed means multi-GPU, likely multi-node, inference infrastructure, plus the engineering time to manage sharding, quantization compatibility, and uptime. Compare that to a Claude Opus 5 or Claude Fable 5 API call, where Anthropic owns all of that infrastructure risk and you pay per token with zero hardware to manage.

Self-hosting K3 makes sense when you have a specific reason: data residency requirements that forbid sending prompts to a third-party API, an existing GPU cluster you're already amortizing, or a workload with volume high enough that the infrastructure cost beats API pricing over time. For most teams building an AI product or an internal automation layer, that math still favors an API-first approach, at least until self-hosting costs come down further. It's the same build-versus-buy calculation we walk clients through in AI automation agency engagements, just applied to model infrastructure instead of workflow tooling.

Kimi K3 vs Claude Opus 5 and Claude Fable 5

Here's the part worth being precise about: there is no independently verified, apples-to-apples benchmark of Kimi K3 against Claude Opus 5 as of this writing.

Claude Opus 5 launched July 24, 2026, three days before K3's weights went live, and currently ranks #1 of 170 models on the Artificial Analysis Intelligence Index at a score of 61, ahead of Claude Fable 5 at 60. Moonshot's own K3 model card, published alongside the July 27 release, benchmarks K3 against Claude Opus 4.8 and Claude Fable 5, the prior Anthropic generation and Anthropic's other current flagship, not against Opus 5 directly. That's not a knock on Moonshot; their evaluation window likely closed before Opus 5's benchmarks were public. But it means any headline claiming K3 "beats Opus 5" or vice versa is going further than the actual published data supports right now.

What we can say with confidence:

• Availability: Opus 5 and Fable 5 are closed, API-only models you access through Anthropic. K3 is the only one of the three you can download and run on your own hardware.
• Context window: K3's 1-million-token window matches the top end of what current frontier models offer.
• Cost structure: Opus 5 and Fable 5 bill per token with no infrastructure to manage. K3 is free to download but shifts the cost to whatever infrastructure you build to run it, or to a hosted K3 endpoint from a third party.
• Benchmarks against the actual current Anthropic flagship: not yet published by either side. Treat any specific comparative score you see right now as unverified until someone runs a controlled, independent test.

For the full picture on where Opus 5 and Fable 5 stand today, see our Claude Opus 5 benchmarks breakdown and Opus 5 vs Fable 5 comparison.

Should Your Team Actually Self-Host K3?

Run through this before committing:

• Do you have a data residency or compliance requirement that rules out sending prompts to a third-party API? If yes, self-hosting starts to look necessary, not just cheaper.
• Do you already have, or plan to acquire, multi-GPU infrastructure you'd otherwise leave idle? If the hardware cost is already sunk, K3's per-token cost drops toward zero.
• Is your inference volume high and predictable enough that infrastructure costs beat API billing over a 12-month horizon? Run the math with real numbers, not a rough guess, before deciding.
• Does your product's growth trajectory put you anywhere near the license's revenue or MAU thresholds? If so, loop in legal on the Kimi K3 License now, not after you clear $20 million in monthly revenue.
• Do you need day-one access to whatever Anthropic ships next? Opus 5 and Fable 5 update on Anthropic's schedule with no engineering lift on your end. A self-hosted K3 deployment updates on your schedule and your team's effort.

If most of your answers point toward "we just need reliable output at a predictable per-token cost," an API-first setup with Claude still wins on total effort. If you have a specific infrastructure or compliance reason to self-host, K3's release is the first time that's genuinely on the table at this parameter scale. Our AI automation agency work starts with exactly this kind of build-versus-buy assessment before any code gets written.

FAQ

Is Kimi K3 open source? No. Moonshot AI consistently calls it "open weight," not open source, and the Kimi K3 License adds commercial restrictions, including a mandatory separate agreement for large Model-as-a-Service resellers, that a standard open-source license would not include.

When did Kimi K3's open weights release? July 27, 2026, on Hugging Face. Moonshot published the full weight files, model card, and license text at huggingface.co/moonshotai/Kimi-K3 on that date.

Can I use Kimi K3 commercially for free? Yes, for the large majority of use cases. The license only requires a separate paid agreement with Moonshot if you operate a Model-as-a-Service business (reselling API-style access to the model) with aggregate revenue over $20 million across any rolling 12 months.

How many parameters does Kimi K3 have? 2.8 trillion total parameters, with 104 billion activated per token through a mixture-of-experts architecture that selects 16 of 896 experts, per Moonshot's official model card.

Does Kimi K3 beat Claude Opus 5? There's no independently verified benchmark comparing the two as of this writing. Moonshot's own model card compares K3 to Claude Opus 4.8 and Claude Fable 5, not Opus 5, which launched three days before K3's weights went public.

Do I need to display attribution if I build on Kimi K3? Only if your product exceeds 100 million monthly active users or $20 million in monthly revenue. Below that, there's no attribution requirement under the Kimi K3 License.

Is Kimi K3 self-hostable on a single server? Not practically. The full weight repository is roughly 1.56 terabytes, which requires multi-GPU, likely multi-node, infrastructure to run at production speed.

What's the difference between Kimi K3 and Kimi K2.5? K2.5 was Moonshot's previous open-weight release. K3 is the new flagship: larger, natively multimodal, and built on a new attention architecture, but it arrived under a more restrictive commercial license than a plain MIT release.

---

Sources: Moonshot AI, Kimi K3 model card (Hugging Face), Kimi K3 License (Hugging Face), Simon Willison, Kimi K3 release notes, VentureBeat, Kimi K3's full weights are here, Artificial Analysis Intelligence Index