Kimi K3 puts 2.8 trillion parameters behind open weights

Moonshot released Kimi K3 across its app, web, Work, Code, Agent, Agent Swarm and API products on July 16. Full model weights followed on July 27. The staggered release matters because product access and downloadable weights were separate milestones, not one simultaneous event.

The technical specification lists 2.8 trillion total parameters and 104 billion active parameters. It also describes native image input and a context window of 1,048,576 tokens. Those figures make K3 an unusually large open-weight agent model, but they do not make it practical to run on ordinary consumer hardware simply because the files can be downloaded.

Moonshot released the weights under its custom Kimi K3 License. The terms are permissive in several respects but add conditions for large model-as-a-service operators and very large commercial products. The accurate label is open-weight unless a specific legal analysis establishes that a proposed use meets a different definition.

Moonshot's benchmark tables compare K3 with leading proprietary models, although the source ledger notes differences in test harnesses and sources. The results should remain attributed to Moonshot and examined task by task. A parameter count, active parameter count and context window also describe different properties and should not be combined into one general measure of intelligence.

The practical audience is likely to include organizations with substantial serving infrastructure, researchers adapting the weights and developers using the hosted products or API. Future deployment reports should disclose hardware, precision, throughput and license obligations so that readers can distinguish downloadable access from realistic operating cost.