Moonshot AI releases the 2.8-trillion-parameter Kimi K3

Background

Moonshot AI is a Beijing-based company that develops the Kimi assistant and the Kimi family of large language models. In its own reporting, Moonshot compared the new model with its earlier K2.6 release, so K3 is best read as the next generation in that line.

What happened

Moonshot AI released Kimi K3 on 16 July 2026. The model has 2.8 trillion total parameters, and about 104 billion of them are activated for each token it processes.^[3][1] Wikipedia describes the family as the largest open-weights model family.^[1]

A parameter is one of the numbers a model learns during training. Total parameters count all of them, while activated parameters count only the ones used to process a single token. The gap between the two numbers is why a model of this size can cost less per token to run than its total size suggests.

Architecture and context

Attention is the part of a model that decides how strongly each token looks at other tokens. Standard full attention compares each token with every earlier token, so its cost grows quickly as the text gets longer. Linear attention is a family of designs that reduces that cost. In K3, 69 Kimi Delta Attention (KDA) linear-attention layers are mixed with 24 Gated MLA layers.^[3] The context window is 1,048,576 tokens, which is the amount of text the model can consider at one time.^[3]

Benchmark results

Moonshot reported self-reported scores at maximum reasoning effort: GPQA Diamond 93.5, HLE-Full 43.5 without tools and 56.0 with tools, Terminal-Bench 2.1 88.3, BrowseComp 91.2 and OSWorld-Verified 84.8.^[3] Self-reported means the developer ran the tests itself and published the numbers, so independent labs may measure differently. GPQA Diamond is a set of graduate-level science questions. HLE is a hard exam across many subjects. Terminal-Bench measures tasks carried out in a command line, BrowseComp measures finding information on the web, and OSWorld measures tasks on a computer desktop.

Independent rankings

K3 debuted at No. 3 on the Artificial Analysis leaderboard and topped Arena.ai's front-end web development benchmark.^[2][1] Moonshot also reported a private long-horizon Elo rating of 1547 for the model, 732 points above K2.6.^[2][1] Elo is a rating method that turns head-to-head results into a single score, so the figure compares K3 with an earlier model on Moonshot's own private test.

Weights and license

On 27 July 2026 Moonshot released the weights under the Kimi K3 License.^[2][1][3] Open weights means the trained numbers can be downloaded, so others can run and study the model, while a license sets the terms of use. Under this license, hosts earning more than US$20 million a year from K3 must negotiate with Moonshot, and Wikipedia reports revenue sharing of up to 30 percent.^[2][1][3] Large products must also attribute Kimi K3.^[2][1][3]

Significance

For readers, K3 shows how a model with 2.8 trillion parameters can be released with downloadable weights while the license adds conditions for large commercial hosts. The revenue and attribution conditions apply to hosts and large products built on the model, so they matter for anyone deploying it.