Moonshot AI appears poised to release Kimi K3, its most ambitious model yet, on July 15 at midnight China time. A promotional page that briefly appeared on Moonshot's official Kimi API Platform before being replaced with a redirect revealed the launch date and hinted at significant architectural changes from the company's current flagship models.

What's Different About Kimi K3?

The leaked announcement titled "Kimi K3 launch limited-time recharge campaign" suggests this isn't simply a scaled-up version of Moonshot's existing technology. Rather than continuing the Mixture-of-Experts design that powers Kimi K2.6 and K2.7 Code, K3 reportedly features a completely new architecture specifically built for long-horizon agent tasks. This represents a meaningful strategic shift, moving beyond single-turn responses toward multi-step, multi-hour workflows that require sustained reasoning and planning.

Unconfirmed reports point to roughly 2.5 trillion parameters and a 1-million-token context window, though Moonshot AI has not officially published benchmarks or a model card to verify these claims. To put the context window in perspective, a 1-million-token capacity means K3 could process roughly 750,000 words at once, far exceeding most competing models currently available.

How Does This Fit Into the Broader AI Race?

Moonshot's reported pivot toward agent-focused design mirrors competitive moves by other leading AI labs. Anthropic has been pushing aggressively into agentic coding capabilities, and this architectural shift suggests Moonshot is chasing the same frontier rather than simply adding more parameters to an existing formula. The timing places K3 squarely in the middle of one of the most competitive stretches the model market has experienced all year.

The current Kimi K2 series has succeeded partly by keeping inference costs low while maintaining competitive performance. The Mixture-of-Experts design allows the model to scale parameter count without proportionally increasing computational demands, which is why K2.6 and K2.7 Code have undercut rivals like Claude and GPT on price-per-token metrics. A shift toward long-horizon agent tasks will likely show up in how K3 handles complex, multi-step workflows rather than in traditional single-turn benchmarks.

What We Know and Don't Know About the Launch

• Confirmed Details: A promotional page from Moonshot's own API Platform briefly went live, stating the K3 launch campaign begins July 15 at 00:00 China time before being replaced with a redirect.
• Rumored Specifications: Reports suggest approximately 2.5 trillion parameters and a 1-million-token context window, but no official model card or benchmark results have been published.
• Architecture Claims: The leaked page and separate rumors point to a completely new design focused on long-horizon agent tasks, departing from the Mixture-of-Experts approach used in current models.
• Official Confirmation: Moonshot AI has not yet publicly confirmed the launch date, the promotional campaign, or any of the architectural claims surrounding K3.

The precision of the dated language in the leaked promotional copy, which specified an exact launch time rather than vague future planning language, suggests the announcement came from near-final launch materials rather than speculative internal documents. However, leaked pages have a history of being pulled before full details become public, so the real specifications may not settle the architectural questions until an official announcement arrives.

If Kimi K3 does launch as the leak suggests, the model will arrive at a moment when long-context AI development and agent-focused reasoning have become central battlegrounds in the AI industry. The race to build models capable of sustained, multi-step reasoning over extended contexts has been a running theme across the broader AI landscape, with competing labs pushing past the million-token mark and exploring new architectural approaches to handle complex agent workflows.