# A Leaked Kimi K3 Promo Page Shows Moonshot Chasing an Edge Claude Already Has > Source: > Published: 2026-07-16 11:15:28+00:00 *Moonshot AI's next flagship model leaked through its own promo page before anyone at the company said a word about it.* On July 14, a page titled "Kimi K3 launch limited-time recharge campaign" briefly went live on Moonshot's Kimi Open Platform, according to reporting compiled by Essa Mamdani and other outlets tracking the leak. It promised bonus credits tied to a K3 launch starting July 15 at midnight China time. Within hours, Moonshot pulled it and swapped in a redirect. Screenshots had already spread. That's the leak, in full. As of July 16, there is still no model card, no published weights, and no official benchmark sheet. Soft access reportedly appears live on kimi.com, but Moonshot itself has said nothing on the record. ## What the leaked page actually promised The pulled page pointed to something specific: a Mixture-of-Experts model with roughly 2.5 trillion total parameters and a 1 million token context window, aimed squarely at long-horizon coding and agentic workloads, per pre-release leaks reported by Gate News and KuCoin. That's the same territory Anthropic has staked out with Claude. It's not a coincidence. Moonshot is reportedly using a technique it calls Kimi-Linear, a hybrid attention architecture with selective state compression, to make that context window usable rather than theoretical. The claim, per early technical writeups, is a 75% cut in KV cache size and 6x faster decoding at the 1 million token mark. That's a big claim. If it holds up under independent testing, it solves the real problem with long-context models: serving them affordably. Most labs can advertise a huge window. Few can run it without burning through GPU memory. ## The comparison Moonshot doesn't want you to make Here's what undercuts the leap-frog narrative. Claude Opus 4.8 and Sonnet 5 already ship with a 1 million token context window, generally available at standard API pricing since March 2026, when Anthropic dropped its long-context surcharge entirely. A 900,000-token call now costs the same per token as a 9,000-token one. So K3's headline spec isn't new ground. It's Moonshot catching up to a number Anthropic already made routine. Where the real contest sits is coding performance, and the picture there is mixed. Kimi's current model, K2.7 Code, hit 81.1% on MCPMark Verified, ahead of Opus 4.8's 76.4% on that same agentic tool-use benchmark. But on SWE-bench Verified, the industry standard for real-world coding tasks, Opus 4.8 posts 88.6% against numbers Moonshot has not yet matched or even independently published for K2.7. Practitioners running K2.7 against production repositories have told outlets like Kingy AI that the public benchmark claims don't fully hold up outside curated test sets. K3 is supposed to close that gap. Nobody outside Moonshot has seen the numbers yet. ## The distillation accusation hanging over this launch None of this is happening in a vacuum. In February, Anthropic accused three Chinese labs, DeepSeek, MiniMax, and Moonshot, of running a coordinated distillation operation against Claude using roughly 24,000 fraudulent accounts and generating more than 16 million exchanges in violation of its terms of service, according to Anthropic's own disclosure and reporting from VentureBeat. Moonshot's share was over 3.4 million exchanges, the second largest of the three. Anthropic said the queries specifically targeted agentic reasoning and tool use, coding and data analysis, computer-use agent development, and computer vision. Those are the exact categories K3 is now being marketed against. That timeline matters. A lab accused five months ago of mining Claude's own outputs to train its models is now leaking a flagship built for the same workloads Anthropic said it was targeting. Whether K3's architecture is genuinely novel or leans on what Moonshot extracted earlier this year is not something outsiders can verify from a pulled promo page and a teaser video. What is verifiable is the pattern. Chinese labs have moved from trailing frontier US models by a full generation to trailing by months, sometimes weeks. K2.7 already beat Opus 4.8 on one agentic benchmark. If K3 ships with real SWE-bench numbers anywhere close to Anthropic's, the gap in enterprise coding tools, the one product category Anthropic has built its commercial case around, gets a lot harder to defend on capability alone. Moonshot still has to actually publish the model. Until it does, the leaked promo page is the only receipt anyone has. **Also read:** [A Former Red Bull F1 Engineer Just Raised Germany's Biggest Seed Round Ever](https://startupfortune.com/a-former-red-bull-f1-engineer-just-raised-germanys-biggest-seed-round-ever/) • [Jason M Riggs on Why AI Spending Keeps Outrunning AI Leadership](https://startupfortune.com/jason-m-riggs-on-why-ai-spending-keeps-outrunning-ai-leadership/) • [Scribble Network Wants To Close The Gap Between Knowing And Fixing AI Visibility](https://startupfortune.com/scribble-network-wants-to-close-the-gap-between-knowing-and-fixing-ai-visibility/)