The short answer: no — Kimi K3 itself is not down. But if you tried to sign up for Kimi over the weekend of July 19–20, 2026, you probably found the door locked. Demand for Moonshot AI's newly launched model "has pushed close to the limits of our current capacity," the company told users, and so it did something rare for a breakout AI release: it paused new subscriptions.
What Moonshot actually said
In a statement titled To Kimi Users: An Update on Compute Capacity Constraints and Subscription Suspension, Moonshot said that user requests in the 48 hours following the Kimi K3 launch had "surged far beyond projections, bringing its existing compute cluster close to maximum capacity." The fix it chose was blunt: new consumer subscriptions were suspended "with immediate effect."
"To protect the experience of existing subscribers, we're temporarily pausing new subscriptions and prioritising compute for current members," the company said, adding: "Existing subscribers will not be affected." Moonshot said it is "adding capacity as fast as we can and will reopen new subscription spots in batches."
So is it an outage?
Not exactly. The Kimi service itself — the chatbot, Kimi Work, Kimi Code, and the Kimi API — remained available to existing subscribers. What got throttled was onboarding: a deliberate capacity-management decision to protect quality for the people already paying, rather than degrade the product for everyone. Outage-tracking sites logged scattered slow-performance and sign-in complaints, but the model kept running. It's better understood as a sold-out launch than a crash.
Why everyone wanted in: the model behind the surge
The demand that broke Kimi's capacity didn't come from nowhere. Released on July 17, 2026 — the eve of Shanghai's World Artificial Intelligence Conference — Kimi K3 is being treated by much of the industry as the year's defining AI release. "This may be the single biggest release of the year," said Anastasios Angelopoulos, co-founder and CEO of the evaluation platform Arena, "and marks a moment when open-source Chinese models are surpassing closed U.S. models."
By the numbers, K3 is the largest open-weight model ever released: a 2.8-trillion-parameter mixture-of-experts network with a 1-million-token context window, surpassing DeepSeek's V4 Pro (1.6T) and Zhipu's GLM-5 series (744B). It handles text and images natively and is built for long-horizon coding, knowledge work, and multi-step reasoning. On Arena's front-end coding leaderboard — a measure built by UC Berkeley researchers — K3 topped the charts.
Independent benchmarking firm Artificial Analysis gave K3 a score of 57 on its Intelligence Index, ranking it fourth among 189 models tested — behind only Claude Fable 5 and two configurations of OpenAI's GPT-5.6 Sol, and ahead of Anthropic's Claude Opus 4.8, OpenAI's GPT-5.5, and GLM-5.2. University of Pennsylvania professor Ethan Mollick called it "really good, closest to the frontier yet." The hype drew instant comparisons to China's market-shaking DeepSeek moment of early 2025.
Price, and a familiar strategic backdrop
K3's API is priced at $3 per million input tokens and $15 per million output tokens — the highest rate yet for a Chinese AI model, yet still roughly half the cost of Anthropic's Opus 4.8 and below the per-task cost of OpenAI's GPT-5.6 Sol. Full open weights are scheduled for release on July 27, 2026.
The open-weight strategy itself is part of the story. Moonshot — founded in 2023 by Carnegie Mellon Ph.D. Yang Zhilin and backed by Alibaba and Tencent — pivoted toward open releases with K2 in 2025 and K2.5 in January 2026. Facing U.S. chip restrictions and unable to match American rivals' capital spending, Chinese labs including DeepSeek, Alibaba, Tencent, and Baidu have turned to open-weight releases to accelerate adoption, attract developers, and set technical standards. Anthropic has separately accused Moonshot, DeepSeek, and MiniMax of "illicit distillation" of Western models — a claim Beijing calls "groundless."
What happens next
Moonshot says it will reopen new subscription spots in batches as capacity comes online, and it is splitting its offering to manage load more precisely: a Kimi Membership for Kimi Web, App, and Work, and a separate Kimi Code Membershipfor coding workflows. The company framed the split as a way to "match compute more precisely and keep the experience stable." The Kimi consumer app, Kimi Work, Kimi Code, and the Kimi API (also reachable through OpenRouter) continue to serve existing users.
For everyone else, the lesson of the weekend is less about a model being "down" and more about a model being too popular. A trillion-parameter class open-weight model that rivals the best closed systems — at roughly half the price — turned out to be exactly the kind of thing the world wants all at once.