China's biggest ever open AI model hit its GPU limit within 48 hours, exposing the strategy's real weak point

[Fox]

Moonshot AI released Kimi K3 on July 16 with almost no fanfare, described by one observer as no keynote, no model card, just a quiet overnight update to kimi.com, echoing the same low key style DeepSeek used to launch R1 back in January 2025. At 2.8 trillion total parameters, K3 is now the largest open weight AI model ever released, roughly 1.75 times the size of DeepSeek's own V4 Pro and dwarfing Zhipu AI's GLM 5 series, and early benchmarks show it performing competitively with the strongest proprietary systems from Anthropic and OpenAI, actually topping the live Arena WebDev leaderboard for front end coding ahead of both Claude Fable 5 and GPT-5.6 Sol

The technical story is about efficiency as much as raw scale. K3 is a sparse mixture of experts model activating just 16 of 896 experts per token, under 2 percent of its total parameters, using two new architectural techniques Moonshot calls Kimi Delta Attention and Attention Residuals to cut inference costs and improve reasoning quality despite the model's enormous total size. Full open weights are scheduled for release on July 27, letting outside researchers independently verify Moonshot's benchmark claims for the first time

But within 48 hours of launch, Kimi K3 hit real GPU capacity limits, forcing a subscription pause that exposed the actual constraint behind China's open weight AI strategy, chip access, not model quality. Even as Moonshot pushes model capability forward, the underlying compute needed to actually serve that model to paying users remains scarce, a direct consequence of ongoing US semiconductor export controls that Chinese firms have worked around through custom hardware and offshore cloud rental arrangements, some of which Congress moved to close through legislation passed back in January

The release also carries an unresolved shadow. Anthropic publicly accused Moonshot, alongside DeepSeek and MiniMax, back in February of running industrial scale distillation attacks against Claude, involving more than 24,000 fraudulent accounts and roughly 16 million exchanges, with about 3.4 million queries specifically attributed to Moonshot. Moonshot has neither confirmed nor denied the allegation, and K3's technical documentation credits its architectural innovations without addressing where its training data actually came from. Even the self hosting option arriving with the July 27 weight release doesn't resolve the underlying legal question, China's National Intelligence Law obligations apply to Moonshot as a company, regardless of whether any specific inference run happens to pass through Chinese infrastructure or not

Kimi K3 Subscription Pause Exposes GPU Crunch Behind Open-Weight Strategywww.techtimes.com

[Linda52]

A capacity crunch within 48 hours of launch is such a telling real world stress test, no amount of clever architecture work solves the problem if there simply aren't enough chips to actually serve the demand you generated