Moonshot AI introduced Kimi K3, a 2.8 trillion-parameter vision–language model. The company made the model available immediately via API and said it will release the model weights by July 27 — if published as promised, those weights would make Kimi K3 the largest known open-weights model to date.
How it is built
Kimi K3 is a mixture-of-experts (MoE) transformer with a total of 2.8 trillion parameters. The model contains 896 experts and activates about 16 experts per token, roughly corresponding to 50 billion active parameters per token. It accepts text, images, and video as input (up to 1 million tokens) and produces text outputs (up to 1 million tokens), processing at about 62.0 tokens per second.
Two architectural changes Moonshot recently published are central to Kimi K3:
-
Kimi Delta Attention (KDA): a linear attention mechanism that keeps a fixed-size memory for long inputs and decides when to overwrite older entries. In Moonshot’s earlier experimental model Kimi Linear, KDA reduced memory use by up to 75% and increased output speed up to 6× at 1 million-token input lengths.
-
Attention Residuals: instead of standard residual summation, this method applies attention across model depth so each layer can selectively draw on earlier layers’ outputs. Kimi K3 uses a block-level Attention Residuals variant; in Moonshot’s experiments a block-level implementation matched standard residual performance while using 20% less training compute.
KDA is used in three out of every four attention layers in Kimi K3. Moonshot reported that KDA and Attention Residuals, together with a sparser MoE design and improved training and data recipes, made training roughly 2.5× more efficient than its predecessor when measuring model improvement per unit of compute. The company plans to publish more technical details in a forthcoming report.
Capabilities and features
Kimi K3 supports adjustable reasoning levels (currently “max” and always-on are available; low and high modes are promised later), tool use, structured outputs, and automatic context caching.
Performance on benchmarks
Independent evaluations place Kimi K3 just behind the top proprietary models while leading all open-weight models in many tests:
-
Artificial Analysis Intelligence Index: Kimi K3 at max reasoning scored 57, trailing only Claude Fable 5 (max reasoning with fallback, 60) and GPT-5.6 Sol (max reasoning, 59). The nearest open-weights competitor, GLM-5.2 at max reasoning, scored 51.
-
AutomationBench-AA: Kimi K3 (max reasoning) led the field with 53% on automated business workflow tests.
-
GDPval-AA v2: Kimi K3 (max reasoning) scored 1,668 Elo, behind Claude Fable 5 (1,760 Elo) and GPT-5.6 Sol (1,743 Elo).
-
Arena.ai Code Arena — WebDev: Kimi K3 debuted at 1,679 Elo and placed first in six of seven frontend domains.
Cost-effectiveness: on the Artificial Analysis Intelligence Index tasks, Kimi K3 at max reasoning completed tasks at $0.95 per task, comparable to GPT-5.6 Sol ($1.04 per task), cheaper than Claude Fable 5 with fallback ($2.75 per task), and more expensive than GLM-5.2 ($0.47 per task).
Availability and pricing
Kimi K3 is available through Kimi.com, Kimi mobile apps, Kimi Work app, and the Kimi Code CLI. Token pricing is $3.00 / $0.30 / $15.00 per million for input / cached / output tokens, respectively. Kimi app memberships range from a free tier up to $199 per month.
The company has promised to release model weights soon (by July 27), but has not disclosed the active parameter count, the training datasets and methods, or the license under which the weights will be released.
Context and why it matters
Over the past year open models have rapidly iterated and often leapfrogged one another. Moonshot’s Kimi family progressed from Kimi K2 (1 trillion parameters) to Kimi K2.5 and Kimi K2.6, each taking turns at the top of open-weight leaderboards. More recently GLM-5.2 led briefly, and three days after Kimi K3’s launch Alibaba — a Moonshot backer — introduced Qwen3.8-Max-Preview, a 2.4 trillion-parameter preview that the company says trails only Claude Fable 5 and will be released with open weights.
Kimi K3 narrows reasons to default to top proprietary models: it offers near‑state-of-the-art performance at a lower per-task cost in many benchmarks, and, if its weights are released under a permissive license, enables fine-tuning, distillation, and independent analysis. The owner of the released weights will determine usage policies, so open weights can change the practical constraints developers face compared with closed models.
Final note
Moonshot attributes Kimi K3’s efficiency gains to architectural innovations (Kimi Delta Attention, Attention Residuals), increased sparsity and refined training recipes. Two of the core ideas have been published with code; Moonshot says it will publish additional technical material to allow the community to inspect and build on these techniques.
