China just handed the world its most powerful free AI model, and the timing was no accident. Moonshot AI released Kimi K3 on Thursday, carrying 2.8 trillion parameters, which the Beijing lab says makes it the largest open-source model ever built. It landed days before Shanghai hosts the World Artificial Intelligence Conference.
K3 runs roughly 75% larger than DeepSeek’s V4 Pro at 1.6 trillion parameters, and it dwarfs Zhipu’s 744 billion GLM-5 series. Moonshot says this is the ninth time in a year that a Kimi release has set a new open-source size record, a run that started when K2 crossed one trillion.
The performance claims are where this gets interesting, though every number comes from Moonshot itself. The company says K3 substantially outperformed Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.6 Sol, while performing competitively against Claude Fable 5. It concedes the model still trails the strongest proprietary systems overall, which is a refreshingly honest caveat.
K3 costs $15 per million output tokens, against roughly $50 for Fable, so buyers get frontier-adjacent capability at a third of the price. Better still, the full weights go public on July 27, meaning anyone can download and run it free.
The architecture carries real engineering behind it. Moonshot built K3 on Kimi Delta Attention and Attention Residuals, both previously published as open research, alongside a one-million-token context window and native visual understanding. It also works with the OpenAI SDK, which quietly lowers the switching cost for developers already building elsewhere.
For Pakistani developers, that free weight release matters most. Frontier-class capability without dollar subscriptions changes what small teams here can realistically attempt.
