Kimi K3, the next major model from Chinese AI lab Moonshot AI, appears to be launching on July 15, 2025 at midnight China Standard Time — and the announcement came not from a press release, but from a promotional page on Moonshot's own Kimi API platform that briefly went live before being quietly pulled.
How the Kimi K3 Launch Date Leaked
An official-looking page appeared on the Kimi API platform titled "Kimi K3 launch limited-time recharge campaign." The campaign listed a start time of July 15 at 00:00 China Standard Time, strongly implying the model would go live at that moment. Shortly after being spotted, the page began redirecting rather than displaying the announcement — a classic accidental-reveal pattern that AI enthusiasts have learned to read as a near-certain confirmation.
No formal press release, benchmark report, or model card has accompanied the leak. Moonshot AI has not officially commented. But the specificity of the campaign — complete with a recharge promotion tied to a hard launch time — leaves little ambiguity.
What We Know (and Don't Know) About Kimi K3
Details are scarce, but two categories of information have emerged: what Moonshot has previously indicated about K3's design goals, and what separate rumours suggest about its raw scale.
Architecture and Agent Focus
K3 is described as using a completely new architecture, a departure from whatever underpinned Kimi K2 and earlier models. The stated design priority is long-horizon agent tasks — meaning multi-step, goal-directed workflows where a model must plan, execute, and adapt across many actions, rather than simply responding to a single prompt.
This focus aligns with a broader industry shift. Labs like Anthropic with Claude and OpenAI with Codex have been racing to make models genuinely useful as autonomous agents, not just chat assistants. Moonshot appears to be building K3 from the ground up with that use case as the primary target.
Rumoured Scale
- 2.5 trillion parameters — if accurate, this would make K3 one of the largest models ever publicly discussed, dwarfing most frontier models currently available via API.
- 1 million token context window — a context length that would allow K3 to ingest entire codebases, lengthy legal documents, or hours of transcribed audio in a single pass.
It is important to stress that neither figure comes from Moonshot AI directly. These numbers originated in separate, unverified rumours circulating in the AI research community, and they should be treated with appropriate scepticism until the company publishes an official model card or technical report.
Why Kimi K3 Matters for the Frontier AI Race
If even the broad strokes of the K3 rumours prove accurate, the model's arrival would be a significant moment in the ongoing global LLM competition. Chinese labs have already demonstrated they can compete at the frontier — most notably with DeepSeek, whose aggressive pricing has permanently repriced enterprise AI and forced Western competitors to reconsider their cost structures.
A Moonshot model with a trillion-scale parameter count and a native million-token context window would push the industry in two important directions:
- Agent capability benchmarks — Long-horizon reasoning is notoriously hard to evaluate. K3's release would likely come with new benchmarks or dramatically improve existing ones, raising the bar for what developers expect from agentic AI.
- Context window competition — A 1M-token window as a production feature (not just a research demo) would meaningfully expand what developers can build, particularly for enterprise document processing, software engineering agents, and scientific research tools.
This also arrives in a week already packed with AI news. Singapore's Agnes 2.5 Pro recently entered the frontier model race, and OpenAI's Codex platform crossed 7 million active users in the same period — a sign of how quickly the applied AI landscape is moving.
What to Watch For at Launch
When Kimi K3 officially drops — assuming the July 15 date holds — here is what will matter most:
- Official parameter count and architecture details: Whether the 2.5T rumour has any basis, and what the new architecture actually involves (MoE routing? Different attention mechanisms?).
- Benchmark performance: Particularly on agentic and long-context evaluations like SWE-bench, GPQA, and any proprietary agent task benchmarks Moonshot chooses to highlight.
- API pricing and availability: The recharge campaign suggests K3 will be API-accessible from day one, which would make it immediately relevant for developers building agent workflows.
- Context window reality vs. rumour: A 1M-token context is meaningless without strong retrieval and reasoning at that length; real-world tests will matter far more than the headline number.
We will update this article as Moonshot AI publishes official details. The AI space moves fast — and right now, it appears to be moving at midnight on July 15.
