Moonshot AI has released Kimi K3, an open-weight model that is drawing attention not just for its massive size, but for a different kind of bet — one on memory rather than raw computing power. At 2.8 trillion parameters, it is the largest open-weight model ever made publicly available, according to the company’s technical blog.
The model launched on July 16 and immediately stood out because of its parameter count. Model sizes are usually grouped into rough brackets, and 2.8 trillion rounds into what the industry calls the 3T class — a tier no openly available model had entered before. Moonshot’s flagship jumps from just over 1 trillion to 2.8 trillion in a single release, clearing DeepSeek’s 1.6T V4 Pro by a wide margin.
Memory Over Compute: The Real Story Behind Kimi K3
While the parameter count has drawn the most attention, it explains the least about how the model actually runs. According to artificialintelligence-news.com, Moonshot’s Kimi K3 trades compute for memory at nearly every layer. This is a deliberate design choice that shifts the model’s reliance away from heavy computing power and toward memory architecture.
The natural conclusion from this design is that Moonshot has engineered its way around US compute restrictions. Memory is where China’s chip industry is furthest along, and by focusing on memory rather than compute, Moonshot can build a massive model without needing the most advanced US-made chips.
What This Means for China’s AI Industry
This memory-focused approach is significant because it directly addresses the biggest challenge facing Chinese AI companies: limited access to high-end US chips. Instead of trying to match US companies on raw computing power, Moonshot is betting that better memory management can deliver similar results with less compute.
The company’s own technical blog suggests that this architecture allows Kimi K3 to handle large-scale tasks efficiently, even with constrained hardware. This could give Chinese AI developers a practical path forward without relying on cutting-edge US technology.
Our Take: A Smart Workaround, Not a Breakthrough
In our view, Kimi K3 is a clever engineering response to a political problem. Moonshot has not invented a new type of AI — it has found a way to build a very large model using the resources it has available. The memory-first design is practical, but it is not a fundamental advance in AI capability. It is a workaround, and a smart one at that.
For the global AI race, this means China is not giving up on large models. It is simply finding different ways to build them. Whether memory-focused models can match compute-heavy ones in performance remains to be seen, but Kimi K3 shows that the competition is far from over.
