Moonshot's Kimi 3 expected to be China's largest open AI model with 2-3T parameters, rivaling Anthropic's Opus 4.8.

What Happened

The Financial Times reports that Chinese AI startup Moonshot is preparing to release Kimi 3 (Kimi K3), which is projected to be the largest open-weights AI model originating from China. With an estimated parameter count between 2 trillion and 3 trillion, the model is reportedly designed to close the performance gap with Anthropic's anticipated Opus 4.8.

Technical Details

Scaling an open model to the 2-3 trillion parameter regime is a massive engineering feat. Assuming a Mixture-of-Experts (MoE) architecture—which is standard for models of this scale to maintain reasonable inference compute—Kimi 3 will require substantial VRAM for serving, likely pushing the limits of standard multi-node GPU clusters. The sheer scale also raises questions about Moonshot's training infrastructure. Training a 3T parameter model requires massive compute density and interconnect bandwidth, indicating that Chinese AI labs are successfully navigating or bypassing US export restrictions on high-end hardware like the Nvidia H100.

Why It Matters

From a systems and deployment perspective, an open-weights model of this size fundamentally alters the global AI landscape. If Kimi 3 genuinely rivals Anthropic's Opus 4.8, it proves that algorithmic efficiency and alternative hardware orchestration can overcome compute embargoes. For the open-source community, having access to a 2T+ parameter model provides a new ceiling for fine-tuning, distillation, and research. It potentially breaks the Western monopoly on frontier-level open models, offering a heavy-weight alternative to Meta's Llama 3 series.

What to Watch Next

Engineers and researchers should monitor the actual release for three critical factors: the exact licensing terms (whether it is truly open-source or just open-weights with commercial restrictions), the architecture details (specifically the active parameters per token in the MoE setup), and the context window capabilities, an area where previous Kimi models have historically excelled. Independent benchmark data against Llama 3 400B+ and Opus will be required to validate these performance claims.