Is Kimi K3 Open Source? The License, the Real Download Size, and Where to Use It (2026)

Quick Answer (verified August 16, 2026): Kimi K3 is open-weight, not open source, and not MIT-licensed — despite widespread reporting that it shipped under a "Modified MIT" licence. The Hugging Face repository carries a custom "Kimi K3 License" that requires a separate agreement with Moonshot AI for any model-as-a-service business above $20 million in trailing-12-month revenue, and prominent "Kimi K3" branding for products above 100 million monthly active users or $20 million in monthly revenue. The published checkpoint is 1.56 TB across 118 files and 2,779,931,837,184 parameters — so "free to download" is true, and "cheap to run" is not. Everything below is read directly from the repository, not from coverage.

What Kimi K3 Actually Is

Moonshot AI launched Kimi K3 as an API on July 16, 2026 and published the open weights on July 26, 2026, roughly a day ahead of its announced July 27 target. It was immediately described as the largest open-weight model ever released, and that claim holds up.

Here is what the repository itself reports, read from Hugging Face's model metadata and the model's own config.json on August 16, 2026:

Two of those lines contradict what was widely reported at launch, so they are worth taking slowly.

The Licence Is Not MIT

Multiple outlets reported that Kimi K3 shipped under a "Modified MIT" licence, reasoning by analogy from Moonshot's earlier K2.7 Code release. The actual LICENSE file in the K3 repository is a custom document titled "Kimi K3 License," and Hugging Face's own metadata records it as license: other, license_name: kimi-k3.

It is genuinely permissive in its default case. The grant covers use, copying, modification, merging, publication, distribution, sublicensing, sale, deployment, fine-tuning, and derivative works. For an individual, a researcher, or a company running it internally, it behaves much like an MIT licence.

The conditions attach at scale, and there are two:

1. The model-as-a-service revenue threshold. The licence defines "Model as a Service" as giving a third party access to inference or fine-tuning in a way that lets them exercise meaningful control over inputs, parameters, or training data. If a licensee operating such a business — together with its affiliates — exceeds $20 million in aggregate revenue over any consecutive 12 months, it must enter into a separate agreement with Moonshot AI before any commercial use.

Note what that clause does not cover: end-user products where model capability is embedded in specific features, and mere relaying of requests to models hosted by others. A SaaS product with an AI feature is not a model-as-a-service business under this definition. An inference provider is.

2. The attribution threshold. Any commercial product or service with more than 100 million monthly active users or more than $20 million in monthly revenue must display "Kimi K3" prominently in its user interface.

Both conditions are waived for purely internal use, and for access through Moonshot's official products or certified inference partners.

None of this makes Kimi K3 a bad deal. It makes the common shorthand wrong: this is open-weight with commercial thresholds, not open source, and not MIT. If you are choosing it as the licence-safe option, read the file rather than the coverage.

"Free to Download" Is Not "Cheap to Run"

The second correction is arithmetic. The published checkpoint is 1,561 GB — about 1.56 TB — across 118 files, and Moonshot ships it in 8-bit compressed tensors, not at full precision. Reporting that circulated a "1.4 TB" figure was describing a further four-bit quantization, not the official download.

That is before any context is loaded. A 1M-token context window on a 93-layer model with 96 attention heads adds substantial KV cache on top of the weights.

The practical consequence: no single GPU loads this model, quantized or not. Self-hosting Kimi K3 means multi-GPU infrastructure and roughly 1.6 TB of fast storage just to hold the checkpoint. Together AI and Modal both announced day-0 hosted access precisely because most teams calling it "free" were never going to run it themselves.

The mixture-of-experts design does soften the compute side — only 16 of 896 experts activate per token, so the active parameter count per forward pass is a small fraction of 2.78 trillion. That helps throughput. It does not help memory, because all 896 experts must be resident to be routable.

So the honest summary is: the weights are free, the licence is permissive below $20M, and the hardware is not free at all. "Open weights" moved the cost from a per-token bill to a capital expense; it did not remove it.

Where Kimi K3 Is Worth Using

1. When Self-Hosting Is a Requirement, Not a Preference

The strongest case for K3 is regulatory or contractual: data that cannot leave your infrastructure. That is a genuine capability no API tier can offer at any price, and K3 is the strongest model available under that constraint.

2. Very Long Context

A 1M-token window puts K3 in a small group. For whole-codebase reasoning, long document sets, or extended agent sessions, it is competitive with closed frontier models.

3. Frontend and Code Generation

K3 was reported at #1 overall in Frontend Code Arena, and #2 on the Vals AI index and #3 on Artificial Analysis's Intelligence Index — behind Claude Fable and GPT-5.6 Sol Max. For an open-weight model, sitting third on a general intelligence index is the actual news.

4. Cost at Very High Volume

If you are already running multi-GPU infrastructure and your token volume is large enough, owning inference beats renting it. Below that threshold, hosted access from a certified partner is cheaper and the licence thresholds do not apply.

Where It Is the Wrong Choice

• Anyone without multi-GPU infrastructure. A hosted API from any frontier vendor will be cheaper and faster than standing up 1.56 TB of weights.
• Specialist ceilings. Claude Fable 5 and GPT-5.6 Sol still lead on the hardest reasoning benchmarks.
• Model-as-a-service businesses over $20M. Not prohibited, but you need the separate agreement first. Budget legal time, not just GPU time.

What This Means for Making Presentations

The open-weight question rarely changes anything for slide generation, and it is worth being clear about why.

Deck quality is set by two separate things: the reasoning that turns a source document into an argument, and the tool that turns that argument into layout, charts, and a working PowerPoint file. A 2.78-trillion-parameter model improves the first and does nothing at all for the second. You can run the largest open-weight model in the world and still get a deck that exports badly.

What model choice genuinely affects is whether the argument holds up: whether contradictory sources get reconciled rather than averaged, and whether the numbers on a chart actually follow from the document they came from.

ChatSlide handles the second half — real data charts, AI editing, and one-click export to PowerPoint, Keynote, or PDF — and lets you pick the model per deck for the first half, currently from Anthropic's Claude Opus and OpenAI's GPT-5.6 family (Luna, Terra, and Sol). A class talk and a board-level analysis do not need the same reasoning depth, and paying for the deepest model on both is waste.

For a like-for-like comparison on slide output rather than coding benchmarks, see our benchmark of the best AI model for presentations.

The Bottom Line

Kimi K3 is the largest open-weight model ever published — 2,779,931,837,184 parameters, 1M context, 1.56 TB across 118 files — and it is not MIT-licensed and not open source. The custom Kimi K3 License is permissive for internal and small-commercial use, but requires a separate agreement with Moonshot AI for model-as-a-service businesses above $20 million in trailing-12-month revenue, and prominent "Kimi K3" branding above 100M MAU or $20M monthly revenue.

Free to download, genuinely strong, and expensive to run. Choose it when self-hosting is a requirement; use a hosted model when it is a preference.

Try ChatSlide free — build a presentation with a top-tier model in under 60 seconds, no GPU required.

---

Parameter counts, expert configuration, context window, file count, repository size, and licence terms were read directly from the moonshotai/Kimi-K3 Hugging Face repository — its model metadata, config.json, and LICENSE file — on August 16, 2026. Release dates and benchmark placements are as reported by Interconnects and contemporaneous coverage, and have not been independently reproduced by ChatSlide. Last updated August 16, 2026.