Chinese Frontier Model Wave: Hunyuan 3.0, GLM-5.2, LongCat-2.0, MiMo-V2-Omni, Kimi K3
Key Questions
What is Tencent's Hunyuan 3.0?
Hunyuan 3.0 is a 295B MoE model under Apache 2.0 license that matches GPT-5.5 on BrowseComp and agentic tasks, integrated into WeChat.
Which model did Z.ai release and what are its strengths?
Z.ai released GLM-5.2, a text-only model strong in coding and agent tasks at one-sixth the cost of competitors.
What are the specs of Meituan's LongCat-2.0?
LongCat-2.0 is a 1.6T MoE model with 1M context, running on domestic ASIC hardware and released under MIT license.
What did Moonshot AI release with Kimi K3?
Moonshot AI released Kimi K3, a 2.8T MoE open-weight model claiming to rival Fable and beat GPT-5.6 Sol on some benchmarks, with premium pricing.
What broader signal do these Chinese models send?
They indicate China's push for AI independence and cost-competitive frontier capabilities.
Tencent Hunyuan 3.0 (295B MoE, Apache 2.0, matches GPT-5.5 on BrowseComp, agentic tasks, integrated into WeChat). Z.ai GLM-5.2 (text-only, strong coding/agent at one-sixth cost). Meituan LongCat-2.0 (1.6T MoE, 1M context, domestic ASIC, MIT license). Xiaomi MiMo-V2-Omni (omni-modal agentic). New: Moonshot AI released Kimi K3 (2.8T MoE, open-weight, claims to rival Fable and beat GPT-5.6 Sol on some benchmarks, premium pricing). These models signal China's AI independence and cost-competitive frontier capability.