Moonshot AI Launches Kimi K3, a 2.5T-Parameter Flagship with 1M Context

Moonshot AI put Kimi K3 live on July 15, following its usual quiet-launch playbook: the model appeared on kimi.com and the API platform before any official model card or benchmark table. K3 is reported to be a Mixture-of-Experts design at roughly 2.5 trillion total parameters — the successor to the K2 line that made Moonshot a fixture of the open-weights coding leaderboards — and Moonshot describes it as a flagship for long-horizon coding and end-to-end knowledge work.

• Native context window: 1,048,576 tokens (1M), confirmed by the API's model listing
• Reasoning is always on — the API exposes a single "max" thinking effort and returns the chain in a separate reasoning_content field
• Multimodal input: the API flags image and video support alongside text
• API pricing is not yet published; a launch top-up promotion (10–30% bonus credits, through August 11) is the only official pricing signal so far, while third-party routers list it around $3 in / $15 out per 1M tokens

The launch caps a fast release cadence for Moonshot — K2.6 and K2.7 Code shipped only months ago and were quickly picked up by GitHub Copilot and the coding-plan gateways. K3 is available on the arena now: it runs the same one-shot, single-file briefs as every other model here, so you can vote on it blind against Claude, GPT, GLM and the rest of the field.