← All articles
2026-07-177 min read

Kimi K3 explained, from release date to price

Kimi K3 in full: announced 16 July 2026, weights on 27 July, 2.8 trillion parameters, a 1M token context, its own licence, and what it draws in credits.

Kimi K3 is Moonshot AI's current flagship: 2.8 trillion parameters, published as open weights, with a million token context. Two months on from the announcement it is still the newest thing Moonshot has shipped, and it is still the model most of the internet describes wrongly, because a lot of launch-week coverage invented product names that never existed.

The dates

Moonshot announced K3 on 16 July 2026, with the weights promised for later that month. The weights went up on Hugging Face on 27 July 2026. A technical report followed on 28 July as a PDF in the GitHub repository, and was revised on 6 August, so if you came here looking for the Kimi K3 paper, that is where it lives.

One correction while we are on the subject of launch coverage. There is no K3 Max and no K3 Swarm Max. Moonshot published exactly one model repository, moonshotai/Kimi-K3, and its API lists exactly one K3 model id. The word "max" that everyone turned into a product name is a value of the reasoning_effort parameter, which is why Moonshot's own benchmark table has a column headed "Kimi K3 (max)". Agent Swarm is a real Moonshot feature, but it is a separate thing from February 2026 and it belongs to the K2.5 era.

What is inside it

Total parameters2.8 trillion
Active per token104 billion
Layers93
Experts896, of which 16 selected and 2 shared per token
Attention69 Kimi Delta Attention layers and 24 gated MLA layers
Vision encoderMoonViT-V2, 401 million parameters
Context1,048,576 tokens
Vocabulary160K
LicenceKimi K3 License

The ratio is what makes a model this size servable at all: 2.8 trillion parameters, of which 104 billion fire for any given token. The whole thing still has to sit in memory, because the router can reach any expert for the next token, but only a small slice does arithmetic.

Kimi Delta Attention is Moonshot's own linear attention variant, and it is the reason the 1M context is not just a spec line. It also broke conventional prefix caching, which Moonshot says it fixed by contributing a KDA-compatible implementation upstream to vLLM.

The licence, which is not MIT

K2 shipped under a modified MIT licence, and a lot of people have assumed K3 did the same. It did not. K3 has a bespoke licence, named "Kimi K3 License", which starts from MIT and adds two conditions.

The first is a model-as-a-service threshold. If you run a model-as-a-service business and your group revenue passes 20 million US dollars over any consecutive 12 months, you have to sign a separate agreement with Moonshot before any commercial use.

The second is attribution. If your commercial product has more than 100 million monthly active users or more than 20 million US dollars in monthly revenue, "Kimi K3" has to be shown prominently in the interface.

Neither applies to internal use, and neither applies when you reach the model through Moonshot's own products or its certified partners. For an individual, a startup or a company using it internally, this is MIT with extra reading. For anyone building a serious inference business on top of it, it is a term sheet.

Images, and video

K3 is natively multimodal rather than text with a bolted-on adapter. The model card lists text and image; Moonshot's own vision guide goes further and says the model understands video content as well. There is no separate vision checkpoint, so the multimodal capability is the same weights you would use for text.

Thinking is always on

This is the part that decides your bill. Moonshot documents K3 as always reasoning: it returns a reasoning_content field on every response and there is no off switch in the official API. What you control is depth, through reasoning_effort, which takes low, high and max, with max as the default. At launch only max existed and the other two arrived later.

If you integrate against Moonshot directly, there is a trap here. K3 was trained with preserved thinking history, so multi-turn conversations and tool loops need the complete assistant message passed back, reasoning_content and tool_calls included. Send only the text content and the model loses the thread of its own reasoning.

Our own deployment behaves slightly differently from the official API, and we measured it rather than assuming: sending the off signal does quiet it, taking a short answer from 34 completion tokens down to 8. So in LU Labs the Think button is a real switch on K3, and on the most expensive model in the catalogue that switch is the largest lever you have.

The benchmarks, and how to read them

These are Moonshot's own numbers from the model card, run at maximum reasoning effort. Where two figures appear, the second is with tool augmentation.

BenchmarkScore
GPQA Diamond93.5
Humanity's Last Exam, full43.5, or 56.0 with tools
Terminal-Bench 2.188.3
FrontierSWE81.2
BrowseComp91.2
OSWorld-Verified84.8
MMMU-Pro81.6, or 83.4 with tools
MathVision94.3, or 97.8 with tools

Read them with the usual discount for vendor benchmarks, and with two specific caveats that Moonshot itself supplies. Several of the comparison columns are not Moonshot's own runs, they are figures cited from third-party evaluations as of 23 July 2026. And several of the coding scores put K3 on Moonshot's own harness while the competitors run on theirs, which is not a like-for-like setup. To Moonshot's credit, the launch post says plainly that overall performance still trails the top closed models. That is a more honest launch note than most.

Can you run Kimi K3 yourself?

No, and the reason is more specific than "it's big".

The published weights are 1.56 TB across 96 safetensors shards. They are natively MXFP4, with MXFP8 activations, produced by quantisation-aware training from the supervised fine-tuning stage onward. That matters: this is not a full-precision model waiting for someone to shrink it. The 4-bit version is the release, so the usual community quantisation trick has already been spent, and a further squeeze costs quality that has not been trained for.

All of it has to sit in accelerator memory before you allocate a single token of that million-token context. Moonshot's own deployment guidance recommends supernode configurations of 64 or more accelerators, and names vLLM, SGLang and TokenSpeed as the engines to use. That is a rack, not a workstation, and no amount of offloading to system RAM turns it into one.

If what you actually want is a model on your own machine, that is a different and entirely reachable goal: our guide to how to run AI locally starts from what your hardware can hold, and the best open-weight LLMs of 2026 covers the ones that fit.

What it costs

Moonshot's own API prices K3 at 3.00 US dollars per million input tokens, 15.00 dollars per million output tokens, and 0.30 dollars per million input tokens on a cache hit. Moonshot claims above 90 percent cache hit rates on coding workloads through its own infrastructure, which is a real discount if your prompts repeat and irrelevant if they do not.

In LU Labs Cloud you hold credits instead, one credit being $0.00001, and the model draws from your wallet at its real rate: 285,000 credits per million input tokens and 1,425,000 per million output. Read the output column, because output is the side that moves, and remember that with reasoning on, the thinking pass is billed as output.

Every plan and every credit pack reaches it, so there is no tier to climb and no upgrade that unlocks it. The details of getting an account pointed at it, in the browser or from your own code, are on Kimi K3 online.

When K3 is the right call

It is the most expensive thing in our catalogue per output token, and the habit to avoid is leaving it selected for everything. Reach for it when the problem is genuinely hard, when the context is genuinely huge, or when the job involves an image, because no price advantage recovers a capability another model does not have.

For everything else there is a cheaper model that will not embarrass itself. The side-by-side with the obvious alternative is in Kimi K3 vs DeepSeek V4, and what landed alongside it in the catalogue is in the launch note for K3 and Ling 3.0 Flash.

Sources