Kimi K3 Leak: Moonshot AI may be about to drop its most ambitious model yet, and the leak trail is fresh. A page that briefly went live on Kimi’s own docs site suggests K3 isn’t months away. It’s landing tomorrow.

Kimi K3 Leak: What the Leak Shows

The slip-up appears to have come straight from Moonshot’s own Kimi API Platform. A page titled “Kimi K3 launch limited-time recharge campaign” briefly went live, stating the promotion begins July 15 at 00:00 China time, before the page was swapped out for a redirect. That kind of precise, dated language is usually reserved for near-final launch copy, not speculative planning documents.

Beyond the date, the leak points to something more significant than a pricing bump. K3 is reportedly built on a completely new architecture aimed at long-horizon agent tasks, a departure from simply scaling up the Mixture-of-Experts design behind Kimi K2.6 and K2.7 Code, Moonshot’s current flagship coding model available via the Kimi API. Separate rumours point to roughly 2.5 trillion parameters and a 1-million-token context window, though no official benchmarks or model card have surfaced to back any of it up yet.

Detail	What’s Reported
Leaked page title	“Kimi K3 launch limited-time recharge campaign”
Campaign start	July 15, 00:00 China time
Page status now	Redirects instead of displaying the announcement
Architecture	Completely new design, focused on long-horizon agents
Rumoured scale	~2.5 trillion parameters
Rumoured context window	1 million tokens
Benchmarks / model card	None published yet
Official confirmation	None yet from Moonshot AI

Why a New Architecture Matters

Kimi’s K2 series has leaned on a Mixture-of-Experts design that keeps inference costs low while scaling total parameter count, part of why K2.6 and K2.7 Code have undercut rivals like Claude and GPT on price-per-token. A shift toward long-horizon agent tasks specifically suggests Moonshot is chasing the same frontier as Anthropic’s agentic coding push, rather than just adding more parameters to the existing formula. That’s a meaningfully different bet, and one that will show up in how K3 handles multi-step, multi-hour agent workflows rather than single-turn benchmarks.

That race has been a running theme in our coverage of the broader AI landscape, including the leaked details around Claude Opus 5 and the mysterious Honeycomb model and the growing competition among long-context AI models pushing past the million-token mark. If K3 ships tomorrow as the leak suggests, it lands right in the middle of one of the most competitive stretches the model market has seen all year.

The Takeaway

Nothing here is official yet. Moonshot AI hasn’t confirmed the launch date, the promotion, or the architecture claims, and leaked docs pages have a habit of getting pulled before the full picture is public. If Kimi K3 does land tomorrow, expect the real specs to settle the “new architecture” question fast.