H: Moonshot AI will release a sub-100B parameter model (Kimi K3-Mini or similar) within 6 months (by Fe
Moonshot AI will release a sub-100B parameter model (Kimi K3-Mini or similar) within 6 months (by February 2027), explicitly optimized for Huawei Ascend inference.
The 1.56T parameter scale is economically unviable for most enterprise use cases, especially on Chinese hardware where Nvidia H100/B200 supply is restricted. Huawei Ascend 910B/920 chips have limited high-bandwidth memory, making large MoE models with many experts impractical. Moonshot AI needs a smaller model to compete with the thriving Chinese open-source ecosystem (Qwen, DeepSeek, Yi). The declining sentiment on K3 creates urgency to ship something more practical.
Track Moonshot AI's official blog, Hugging Face model releases, and Chinese tech media (36Kr, LatePost) for announcement of a smaller model.